mirror of
https://github.com/modelscope/FunASR.git
synced 2026-10-12 02:19:18 +00:00
funasr2
This commit is contained in:
游雁
@@ -1,16 +0,0 @@
|
||||
# Boundary Aware Transducer (BAT) Result
|
||||
|
||||
## Training Config
|
||||
- 8 gpu(Tesla V100)
|
||||
- Feature info: using 80 dims fbank, global cmvn, speed perturb(0.9, 1.0, 1.1), specaugment
|
||||
- Train config: conf/train_conformer_bat.yaml
|
||||
- LM config: LM was not used
|
||||
- Model size: 90M
|
||||
|
||||
## Results (CER)
|
||||
- Decode config: conf/decode_bat_conformer.yaml
|
||||
|
||||
| testset | CER(%) |
|
||||
|:-----------:|:-------:|
|
||||
| dev | 4.56 |
|
||||
| test | 4.97 |
|
||||
@@ -1 +0,0 @@
|
||||
beam_size: 10
|
||||
@@ -1,108 +0,0 @@
|
||||
encoder: chunk_conformer
|
||||
encoder_conf:
|
||||
activation_type: swish
|
||||
positional_dropout_rate: 0.5
|
||||
time_reduction_factor: 2
|
||||
embed_vgg_like: false
|
||||
subsampling_factor: 4
|
||||
linear_units: 2048
|
||||
output_size: 512
|
||||
attention_heads: 8
|
||||
dropout_rate: 0.5
|
||||
positional_dropout_rate: 0.5
|
||||
attention_dropout_rate: 0.5
|
||||
cnn_module_kernel: 15
|
||||
num_blocks: 12
|
||||
|
||||
# decoder related
|
||||
rnnt_decoder: rnnt
|
||||
rnnt_decoder_conf:
|
||||
embed_size: 512
|
||||
hidden_size: 512
|
||||
embed_dropout_rate: 0.5
|
||||
dropout_rate: 0.5
|
||||
use_embed_mask: true
|
||||
|
||||
predictor: bat_predictor
|
||||
predictor_conf:
|
||||
idim: 512
|
||||
threshold: 1.0
|
||||
l_order: 1
|
||||
r_order: 1
|
||||
return_accum: true
|
||||
|
||||
joint_network_conf:
|
||||
joint_space_size: 512
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
|
||||
# Auxiliary CTC
|
||||
model: bat
|
||||
model_conf:
|
||||
auxiliary_ctc_weight: 0.0
|
||||
cif_weight: 1.0
|
||||
r_d: 3
|
||||
r_u: 5
|
||||
|
||||
# minibatch related
|
||||
use_amp: true
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 100
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- loss
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- cer_transducer
|
||||
- min
|
||||
keep_nbest_models: 10
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.001
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 25000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 40
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_range:
|
||||
- 0
|
||||
- 50
|
||||
num_time_mask: 5
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 25000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,210 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
|
||||
gpu_num=8
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=5
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_conformer_bat.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_bat_conformer.yaml
|
||||
inference_asr_model=valid.cer_transducer.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} ${feats_dir}/data/${train_set}
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/char/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/char/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=./ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type char \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/cmvn.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/cmvn.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode bat \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../transformer/utils
|
||||
@@ -1,15 +0,0 @@
|
||||
# Branchformer Result
|
||||
|
||||
## Training Config
|
||||
- Feature info: using raw speech, extracting 80 dims fbank online, global cmvn, speed perturb(0.9, 1.0, 1.1), specaugment
|
||||
- Train info: lr 0.001, batch_size 10000, 4 gpu(Tesla V100), acc_grad 1, 180 epochs
|
||||
- Train config: conf/train_asr_branchformer.yaml
|
||||
- LM config: LM was not used
|
||||
|
||||
## Results (CER)
|
||||
- Decode config: conf/decode_asr_transformer.yaml (ctc weight:0.4)
|
||||
|
||||
| testset | CER(%) |
|
||||
|:-----------:|:-------:|
|
||||
| dev | 4.15 |
|
||||
| test | 4.51 |
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 10
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.4
|
||||
lm_weight: 0.0
|
||||
@@ -1,104 +0,0 @@
|
||||
# network architecture
|
||||
# encoder related
|
||||
encoder: branchformer
|
||||
encoder_conf:
|
||||
output_size: 256
|
||||
use_attn: true
|
||||
attention_heads: 4
|
||||
attention_layer_type: rel_selfattn
|
||||
pos_enc_layer_type: rel_pos
|
||||
rel_pos_type: latest
|
||||
use_cgmlp: true
|
||||
cgmlp_linear_units: 2048
|
||||
cgmlp_conv_kernel: 31
|
||||
use_linear_after_conv: false
|
||||
gate_activation: identity
|
||||
merge_method: concat
|
||||
cgmlp_weight: 0.5 # used only if merge_method is "fixed_ave"
|
||||
attn_branch_drop_rate: 0.0 # used only if merge_method is "learned_ave"
|
||||
num_blocks: 24
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
attention_dropout_rate: 0.1
|
||||
input_layer: conv2d
|
||||
stochastic_depth_rate: 0.0
|
||||
|
||||
# decoder related
|
||||
decoder: transformer
|
||||
decoder_conf:
|
||||
attention_heads: 4
|
||||
linear_units: 2048
|
||||
num_blocks: 6
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
self_attention_dropout_rate: 0.
|
||||
src_attention_dropout_rate: 0.
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
# hybrid CTC/attention
|
||||
model_conf:
|
||||
ctc_weight: 0.3
|
||||
lsm_weight: 0.1 # label smoothing option
|
||||
length_normalized_loss: false
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 180
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- acc
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- acc
|
||||
- max
|
||||
keep_nbest_models: 10
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.001
|
||||
weight_decay: 0.000001
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 35000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 27
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_ratio_range:
|
||||
- 0.
|
||||
- 0.05
|
||||
num_time_mask: 10
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 10000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
normalize: None
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,105 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (author: Daniel Povey)
|
||||
# 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
remove_archive=false
|
||||
|
||||
if [ "$1" == --remove-archive ]; then
|
||||
remove_archive=true
|
||||
shift
|
||||
fi
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
|
||||
echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
|
||||
echo "With --remove-archive it will remove the archive after successfully un-tarring it."
|
||||
echo "<corpus-part> can be one of: data_aishell, resource_aishell."
|
||||
fi
|
||||
|
||||
data=$1
|
||||
url=$2
|
||||
part=$3
|
||||
|
||||
if [ ! -d "$data" ]; then
|
||||
echo "$0: no such directory $data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
part_ok=false
|
||||
list="data_aishell resource_aishell"
|
||||
for x in $list; do
|
||||
if [ "$part" == $x ]; then part_ok=true; fi
|
||||
done
|
||||
if ! $part_ok; then
|
||||
echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -z "$url" ]; then
|
||||
echo "$0: empty URL base."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -f $data/$part/.complete ]; then
|
||||
echo "$0: data part $part was already successfully extracted, nothing to do."
|
||||
exit 0;
|
||||
fi
|
||||
|
||||
# sizes of the archive files in bytes.
|
||||
sizes="15582913665 1246920"
|
||||
|
||||
if [ -f $data/$part.tgz ]; then
|
||||
size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
|
||||
size_ok=false
|
||||
for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
|
||||
if ! $size_ok; then
|
||||
echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
|
||||
echo "does not equal the size of one of the archives."
|
||||
rm $data/$part.tgz
|
||||
else
|
||||
echo "$data/$part.tgz exists and appears to be complete."
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ ! -f $data/$part.tgz ]; then
|
||||
if ! command -v wget >/dev/null; then
|
||||
echo "$0: wget is not installed."
|
||||
exit 1;
|
||||
fi
|
||||
full_url=$url/$part.tgz
|
||||
echo "$0: downloading data from $full_url. This may take some time, please be patient."
|
||||
|
||||
cd $data || exit 1
|
||||
if ! wget --no-check-certificate $full_url; then
|
||||
echo "$0: error executing wget $full_url"
|
||||
exit 1;
|
||||
fi
|
||||
fi
|
||||
|
||||
cd $data || exit 1
|
||||
|
||||
if ! tar -xvzf $part.tgz; then
|
||||
echo "$0: error un-tarring archive $data/$part.tgz"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
touch $data/$part/.complete
|
||||
|
||||
if [ $part == "data_aishell" ]; then
|
||||
cd $data/$part/wav || exit 1
|
||||
for wav in ./*.tar.gz; do
|
||||
echo "Extracting wav from $wav"
|
||||
tar -zxf $wav && rm $wav
|
||||
done
|
||||
fi
|
||||
|
||||
echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
|
||||
|
||||
if $remove_archive; then
|
||||
echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
|
||||
rm $data/$part.tgz
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,225 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1,2,3"
|
||||
gpu_num=4
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=5
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_asr_branchformer.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_asr_transformer.yaml
|
||||
inference_asr_model=valid.acc.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/$token_type/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type $token_type \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode asr \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
|
||||
# Prepare files for ModelScope fine-tuning and inference
|
||||
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
|
||||
echo "stage 6: ModelScope Preparation"
|
||||
cp ${feats_dir}/data/${train_set}/cmvn/am.mvn ${exp_dir}/exp/${model_dir}/am.mvn
|
||||
vocab_size=$(cat ${token_list} | wc -l)
|
||||
python utils/gen_modelscope_configuration.py \
|
||||
--am_model_name $inference_asr_model \
|
||||
--mode asr \
|
||||
--model_name conformer \
|
||||
--dataset aishell \
|
||||
--output_dir $exp_dir/exp/$model_dir \
|
||||
--vocab_size $vocab_size \
|
||||
--tag $tag
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../transformer/utils
|
||||
@@ -1,17 +0,0 @@
|
||||
|
||||
# Conformer Result
|
||||
|
||||
## Training Config
|
||||
- Feature info: using 80 dims fbank, global cmvn, speed perturb(0.9, 1.0, 1.1), specaugment
|
||||
- Train info: lr 5e-4, batch_size 25000, 2 gpu(Tesla V100), acc_grad 1, 50 epochs
|
||||
- Train config: conf/train_asr_transformer.yaml
|
||||
- LM config: LM was not used
|
||||
- Model size: 46M
|
||||
|
||||
## Results (CER)
|
||||
- Decode config: conf/decode_asr_transformer.yaml (ctc weight:0.5)
|
||||
|
||||
| testset | CER(%) |
|
||||
|:-----------:|:-------:|
|
||||
| dev | 4.42 |
|
||||
| test | 4.87 |
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 10
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.5
|
||||
lm_weight: 0.7
|
||||
@@ -1,98 +0,0 @@
|
||||
# network architecture
|
||||
# encoder related
|
||||
encoder: conformer
|
||||
encoder_conf:
|
||||
output_size: 256 # dimension of attention
|
||||
attention_heads: 4
|
||||
linear_units: 2048 # the number of units of position-wise feed forward
|
||||
num_blocks: 12 # the number of encoder blocks
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
attention_dropout_rate: 0.0
|
||||
input_layer: conv2d # encoder architecture type
|
||||
normalize_before: true
|
||||
pos_enc_layer_type: rel_pos
|
||||
selfattention_layer_type: rel_selfattn
|
||||
activation_type: swish
|
||||
macaron_style: true
|
||||
use_cnn_module: true
|
||||
cnn_module_kernel: 15
|
||||
|
||||
# decoder related
|
||||
decoder: transformer
|
||||
decoder_conf:
|
||||
attention_heads: 4
|
||||
linear_units: 2048
|
||||
num_blocks: 6
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
self_attention_dropout_rate: 0.0
|
||||
src_attention_dropout_rate: 0.0
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
# hybrid CTC/attention
|
||||
model_conf:
|
||||
ctc_weight: 0.3
|
||||
lsm_weight: 0.1 # label smoothing option
|
||||
length_normalized_loss: false
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 150
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- acc
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- acc
|
||||
- max
|
||||
keep_nbest_models: 10
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.0005
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 30000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 30
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_range:
|
||||
- 0
|
||||
- 40
|
||||
num_time_mask: 2
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 25000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
normalize: None
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,105 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (author: Daniel Povey)
|
||||
# 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
remove_archive=false
|
||||
|
||||
if [ "$1" == --remove-archive ]; then
|
||||
remove_archive=true
|
||||
shift
|
||||
fi
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
|
||||
echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
|
||||
echo "With --remove-archive it will remove the archive after successfully un-tarring it."
|
||||
echo "<corpus-part> can be one of: data_aishell, resource_aishell."
|
||||
fi
|
||||
|
||||
data=$1
|
||||
url=$2
|
||||
part=$3
|
||||
|
||||
if [ ! -d "$data" ]; then
|
||||
echo "$0: no such directory $data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
part_ok=false
|
||||
list="data_aishell resource_aishell"
|
||||
for x in $list; do
|
||||
if [ "$part" == $x ]; then part_ok=true; fi
|
||||
done
|
||||
if ! $part_ok; then
|
||||
echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -z "$url" ]; then
|
||||
echo "$0: empty URL base."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -f $data/$part/.complete ]; then
|
||||
echo "$0: data part $part was already successfully extracted, nothing to do."
|
||||
exit 0;
|
||||
fi
|
||||
|
||||
# sizes of the archive files in bytes.
|
||||
sizes="15582913665 1246920"
|
||||
|
||||
if [ -f $data/$part.tgz ]; then
|
||||
size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
|
||||
size_ok=false
|
||||
for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
|
||||
if ! $size_ok; then
|
||||
echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
|
||||
echo "does not equal the size of one of the archives."
|
||||
rm $data/$part.tgz
|
||||
else
|
||||
echo "$data/$part.tgz exists and appears to be complete."
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ ! -f $data/$part.tgz ]; then
|
||||
if ! command -v wget >/dev/null; then
|
||||
echo "$0: wget is not installed."
|
||||
exit 1;
|
||||
fi
|
||||
full_url=$url/$part.tgz
|
||||
echo "$0: downloading data from $full_url. This may take some time, please be patient."
|
||||
|
||||
cd $data || exit 1
|
||||
if ! wget --no-check-certificate $full_url; then
|
||||
echo "$0: error executing wget $full_url"
|
||||
exit 1;
|
||||
fi
|
||||
fi
|
||||
|
||||
cd $data || exit 1
|
||||
|
||||
if ! tar -xvzf $part.tgz; then
|
||||
echo "$0: error un-tarring archive $data/$part.tgz"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
touch $data/$part/.complete
|
||||
|
||||
if [ $part == "data_aishell" ]; then
|
||||
cd $data/$part/wav || exit 1
|
||||
for wav in ./*.tar.gz; do
|
||||
echo "Extracting wav from $wav"
|
||||
tar -zxf $wav && rm $wav
|
||||
done
|
||||
fi
|
||||
|
||||
echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
|
||||
|
||||
if $remove_archive; then
|
||||
echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
|
||||
rm $data/$part.tgz
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,225 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1"
|
||||
gpu_num=2
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=5
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_asr_conformer.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_asr_transformer.yaml
|
||||
inference_asr_model=valid.acc.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/$token_type/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type $token_type \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode asr \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
|
||||
# Prepare files for ModelScope fine-tuning and inference
|
||||
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
|
||||
echo "stage 6: ModelScope Preparation"
|
||||
cp ${feats_dir}/data/${train_set}/cmvn/am.mvn ${exp_dir}/exp/${model_dir}/am.mvn
|
||||
vocab_size=$(cat ${token_list} | wc -l)
|
||||
python utils/gen_modelscope_configuration.py \
|
||||
--am_model_name $inference_asr_model \
|
||||
--mode asr \
|
||||
--model_name conformer \
|
||||
--dataset aishell \
|
||||
--output_dir $exp_dir/exp/$model_dir \
|
||||
--vocab_size $vocab_size \
|
||||
--tag $tag
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../transformer/utils/
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 1
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.0
|
||||
lm_weight: 0.15
|
||||
-121
@@ -1,121 +0,0 @@
|
||||
# network architecture
|
||||
# encoder related
|
||||
encoder: data2vec_encoder
|
||||
encoder_conf:
|
||||
extractor_mode: layer_norm
|
||||
encoder_layerdrop: 0.1
|
||||
dropout_input: 0.0
|
||||
dropout_features: 0.0
|
||||
feature_grad_mult: 0.0
|
||||
encoder_embed_dim: 768
|
||||
|
||||
mask_prob: 0.65
|
||||
mask_length: 10
|
||||
|
||||
loss_beta: 0
|
||||
loss_scale: null
|
||||
|
||||
instance_norm_target_layer: true
|
||||
average_top_k_layers: 8
|
||||
|
||||
pos_conv_depth: 5
|
||||
conv_pos: 95
|
||||
|
||||
ema_decay: 0.999
|
||||
ema_end_decay: 0.9999
|
||||
ema_anneal_end_step: 30000
|
||||
ema_transformer_only: true
|
||||
ema_layers_only: true
|
||||
|
||||
require_same_masks: true
|
||||
mask_dropout: 0
|
||||
|
||||
# decoder related
|
||||
decoder: paraformer_decoder_san
|
||||
decoder_conf:
|
||||
attention_heads: 12
|
||||
linear_units: 3072
|
||||
num_blocks: 6
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
self_attention_dropout_rate: 0.0
|
||||
src_attention_dropout_rate: 0.0
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
# hybrid CTC/attention
|
||||
model: paraformer
|
||||
model_conf:
|
||||
ctc_weight: 0.3
|
||||
lsm_weight: 0.1
|
||||
length_normalized_loss: false
|
||||
predictor_weight: 1.0
|
||||
sampling_ratio: 0.4
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 150
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- acc
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- acc
|
||||
- max
|
||||
keep_nbest_models: 10
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.00002
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 30000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 30
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_range:
|
||||
- 0
|
||||
- 40
|
||||
num_time_mask: 2
|
||||
|
||||
predictor: cif_predictor
|
||||
predictor_conf:
|
||||
idim: 768
|
||||
threshold: 1.0
|
||||
l_order: 1
|
||||
r_order: 1
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 25000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
unused_parameters: true
|
||||
normalize: None
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,105 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (author: Daniel Povey)
|
||||
# 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
remove_archive=false
|
||||
|
||||
if [ "$1" == --remove-archive ]; then
|
||||
remove_archive=true
|
||||
shift
|
||||
fi
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
|
||||
echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
|
||||
echo "With --remove-archive it will remove the archive after successfully un-tarring it."
|
||||
echo "<corpus-part> can be one of: data_aishell, resource_aishell."
|
||||
fi
|
||||
|
||||
data=$1
|
||||
url=$2
|
||||
part=$3
|
||||
|
||||
if [ ! -d "$data" ]; then
|
||||
echo "$0: no such directory $data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
part_ok=false
|
||||
list="data_aishell resource_aishell"
|
||||
for x in $list; do
|
||||
if [ "$part" == $x ]; then part_ok=true; fi
|
||||
done
|
||||
if ! $part_ok; then
|
||||
echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -z "$url" ]; then
|
||||
echo "$0: empty URL base."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -f $data/$part/.complete ]; then
|
||||
echo "$0: data part $part was already successfully extracted, nothing to do."
|
||||
exit 0;
|
||||
fi
|
||||
|
||||
# sizes of the archive files in bytes.
|
||||
sizes="15582913665 1246920"
|
||||
|
||||
if [ -f $data/$part.tgz ]; then
|
||||
size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
|
||||
size_ok=false
|
||||
for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
|
||||
if ! $size_ok; then
|
||||
echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
|
||||
echo "does not equal the size of one of the archives."
|
||||
rm $data/$part.tgz
|
||||
else
|
||||
echo "$data/$part.tgz exists and appears to be complete."
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ ! -f $data/$part.tgz ]; then
|
||||
if ! command -v wget >/dev/null; then
|
||||
echo "$0: wget is not installed."
|
||||
exit 1;
|
||||
fi
|
||||
full_url=$url/$part.tgz
|
||||
echo "$0: downloading data from $full_url. This may take some time, please be patient."
|
||||
|
||||
cd $data || exit 1
|
||||
if ! wget --no-check-certificate $full_url; then
|
||||
echo "$0: error executing wget $full_url"
|
||||
exit 1;
|
||||
fi
|
||||
fi
|
||||
|
||||
cd $data || exit 1
|
||||
|
||||
if ! tar -xvzf $part.tgz; then
|
||||
echo "$0: error un-tarring archive $data/$part.tgz"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
touch $data/$part/.complete
|
||||
|
||||
if [ $part == "data_aishell" ]; then
|
||||
cd $data/$part/wav || exit 1
|
||||
for wav in ./*.tar.gz; do
|
||||
echo "Extracting wav from $wav"
|
||||
tar -zxf $wav && rm $wav
|
||||
done
|
||||
fi
|
||||
|
||||
echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
|
||||
|
||||
if $remove_archive; then
|
||||
echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
|
||||
rm $data/$part.tgz
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,231 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1"
|
||||
gpu_num=2
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=1
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
model_name=damo/speech_data2vec_pretrain-zh-cn-aishell2-16k-pytorch
|
||||
init_param="$HOME/.cache/modelscope/hub/$model_name/basemodel.pb"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_asr_paraformer_transformer_12e_6d_3072_768.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_asr_transformer_noctc_1best.yaml
|
||||
inference_asr_model=valid.acc.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/$token_type/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
python utils/download_model.py --model_name ${model_name} # download pretrained model on ModelScope
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type $token_type \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--init_param ${init_param} \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode paraformer \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
|
||||
# Prepare files for ModelScope fine-tuning and inference
|
||||
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
|
||||
echo "stage 6: ModelScope Preparation"
|
||||
cp ${feats_dir}/data/${train_set}/cmvn/am.mvn ${exp_dir}/exp/${model_dir}/am.mvn
|
||||
vocab_size=$(cat ${token_list} | wc -l)
|
||||
python utils/gen_modelscope_configuration.py \
|
||||
--am_model_name $inference_asr_model \
|
||||
--mode paraformer \
|
||||
--model_name data2vec_finetune_paraformer \
|
||||
--dataset aishell \
|
||||
--output_dir $exp_dir/exp/$model_dir \
|
||||
--vocab_size $vocab_size \
|
||||
--nat _nat \
|
||||
--tag $tag
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../../aishell/transformer/utils
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 10
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 1.0
|
||||
lm_weight: 0.7
|
||||
-112
@@ -1,112 +0,0 @@
|
||||
# network architecture
|
||||
# encoder related
|
||||
encoder: data2vec_encoder
|
||||
encoder_conf:
|
||||
extractor_mode: layer_norm
|
||||
encoder_layerdrop: 0.1
|
||||
dropout_input: 0.0
|
||||
dropout_features: 0.0
|
||||
feature_grad_mult: 0.0
|
||||
encoder_embed_dim: 768
|
||||
|
||||
mask_prob: 0.65
|
||||
mask_length: 10
|
||||
|
||||
loss_beta: 0
|
||||
loss_scale: null
|
||||
|
||||
instance_norm_target_layer: true
|
||||
average_top_k_layers: 8
|
||||
|
||||
pos_conv_depth: 5
|
||||
conv_pos: 95
|
||||
|
||||
ema_decay: 0.999
|
||||
ema_end_decay: 0.9999
|
||||
ema_anneal_end_step: 30000
|
||||
ema_transformer_only: true
|
||||
ema_layers_only: true
|
||||
|
||||
require_same_masks: true
|
||||
mask_dropout: 0
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
# hybrid CTC/attention
|
||||
model_conf:
|
||||
ctc_weight: 1.0
|
||||
lsm_weight: 0.1 # label smoothing option
|
||||
length_normalized_loss: false
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
patience: none
|
||||
max_epoch: 150
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- acc
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- cer_ctc
|
||||
- min
|
||||
keep_nbest_models: 10
|
||||
|
||||
# NoamLR is deprecated. Use WarmupLR.
|
||||
# The following is equivalent setting for NoamLR:
|
||||
#
|
||||
# optim: adam
|
||||
# optim_conf:
|
||||
# lr: 10.
|
||||
# scheduler: noamlr
|
||||
# scheduler_conf:
|
||||
# model_size: 256
|
||||
# warmup_steps: 25000
|
||||
#
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.00005
|
||||
scheduler: warmuplr # pytorch v1.1.0+ required
|
||||
scheduler_conf:
|
||||
warmup_steps: 25000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 30
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_range:
|
||||
- 0
|
||||
- 40
|
||||
num_time_mask: 2
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 25000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
unused_parameters: true
|
||||
normalize: None
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,105 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (author: Daniel Povey)
|
||||
# 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
remove_archive=false
|
||||
|
||||
if [ "$1" == --remove-archive ]; then
|
||||
remove_archive=true
|
||||
shift
|
||||
fi
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
|
||||
echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
|
||||
echo "With --remove-archive it will remove the archive after successfully un-tarring it."
|
||||
echo "<corpus-part> can be one of: data_aishell, resource_aishell."
|
||||
fi
|
||||
|
||||
data=$1
|
||||
url=$2
|
||||
part=$3
|
||||
|
||||
if [ ! -d "$data" ]; then
|
||||
echo "$0: no such directory $data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
part_ok=false
|
||||
list="data_aishell resource_aishell"
|
||||
for x in $list; do
|
||||
if [ "$part" == $x ]; then part_ok=true; fi
|
||||
done
|
||||
if ! $part_ok; then
|
||||
echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -z "$url" ]; then
|
||||
echo "$0: empty URL base."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -f $data/$part/.complete ]; then
|
||||
echo "$0: data part $part was already successfully extracted, nothing to do."
|
||||
exit 0;
|
||||
fi
|
||||
|
||||
# sizes of the archive files in bytes.
|
||||
sizes="15582913665 1246920"
|
||||
|
||||
if [ -f $data/$part.tgz ]; then
|
||||
size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
|
||||
size_ok=false
|
||||
for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
|
||||
if ! $size_ok; then
|
||||
echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
|
||||
echo "does not equal the size of one of the archives."
|
||||
rm $data/$part.tgz
|
||||
else
|
||||
echo "$data/$part.tgz exists and appears to be complete."
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ ! -f $data/$part.tgz ]; then
|
||||
if ! command -v wget >/dev/null; then
|
||||
echo "$0: wget is not installed."
|
||||
exit 1;
|
||||
fi
|
||||
full_url=$url/$part.tgz
|
||||
echo "$0: downloading data from $full_url. This may take some time, please be patient."
|
||||
|
||||
cd $data || exit 1
|
||||
if ! wget --no-check-certificate $full_url; then
|
||||
echo "$0: error executing wget $full_url"
|
||||
exit 1;
|
||||
fi
|
||||
fi
|
||||
|
||||
cd $data || exit 1
|
||||
|
||||
if ! tar -xvzf $part.tgz; then
|
||||
echo "$0: error un-tarring archive $data/$part.tgz"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
touch $data/$part/.complete
|
||||
|
||||
if [ $part == "data_aishell" ]; then
|
||||
cd $data/$part/wav || exit 1
|
||||
for wav in ./*.tar.gz; do
|
||||
echo "Extracting wav from $wav"
|
||||
tar -zxf $wav && rm $wav
|
||||
done
|
||||
fi
|
||||
|
||||
echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
|
||||
|
||||
if $remove_archive; then
|
||||
echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
|
||||
rm $data/$part.tgz
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,230 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1"
|
||||
gpu_num=2
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=5
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
model_name=damo/speech_data2vec_pretrain-zh-cn-aishell2-16k-pytorch
|
||||
init_param="$HOME/.cache/modelscope/hub/$model_name/basemodel.pb"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_asr_transformer_12e_6d_3072_768.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_asr_transformer.yaml
|
||||
inference_asr_model=valid.cer_ctc.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/$token_type/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
python utils/download_model.py --model_name ${model_name} # download pretrained model on ModelScope
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type $token_type \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--init_param ${init_param} \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode asr \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
|
||||
# Prepare files for ModelScope fine-tuning and inference
|
||||
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
|
||||
echo "stage 6: ModelScope Preparation"
|
||||
cp ${feats_dir}/data/${train_set}/cmvn/am.mvn ${exp_dir}/exp/${model_dir}/am.mvn
|
||||
vocab_size=$(cat ${token_list} | wc -l)
|
||||
python utils/gen_modelscope_configuration.py \
|
||||
--am_model_name $inference_asr_model \
|
||||
--mode asr \
|
||||
--model_name data2vec_finetune_transformer \
|
||||
--dataset aishell \
|
||||
--output_dir $exp_dir/exp/$model_dir \
|
||||
--vocab_size $vocab_size \
|
||||
--tag $tag
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../transformer/utils/
|
||||
@@ -1,15 +0,0 @@
|
||||
# E-Branchformer Result
|
||||
|
||||
## Training Config
|
||||
- Feature info: using raw speech, extracting 80 dims fbank online, global cmvn, speed perturb(0.9, 1.0, 1.1), specaugment
|
||||
- Train info: lr 0.001, batch_size 10000, 4 gpu(Tesla V100), acc_grad 1, 180 epochs
|
||||
- Train config: conf/train_asr_e_branchformer.yaml
|
||||
- LM config: LM was not used
|
||||
|
||||
## Results (CER)
|
||||
- Decode config: conf/decode_asr_transformer.yaml (ctc weight:0.4)
|
||||
|
||||
| testset | CER(%) |
|
||||
|:-----------:|:-------:|
|
||||
| dev | 4.10 |
|
||||
| test | 4.52 |
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 10
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.4
|
||||
lm_weight: 0.0
|
||||
@@ -1,101 +0,0 @@
|
||||
# network architecture
|
||||
# encoder related
|
||||
encoder: e_branchformer
|
||||
encoder_conf:
|
||||
output_size: 256
|
||||
attention_heads: 4
|
||||
attention_layer_type: rel_selfattn
|
||||
pos_enc_layer_type: rel_pos
|
||||
rel_pos_type: latest
|
||||
cgmlp_linear_units: 1024
|
||||
cgmlp_conv_kernel: 31
|
||||
use_linear_after_conv: false
|
||||
gate_activation: identity
|
||||
num_blocks: 12
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
attention_dropout_rate: 0.1
|
||||
input_layer: conv2d
|
||||
layer_drop_rate: 0.0
|
||||
linear_units: 1024
|
||||
positionwise_layer_type: linear
|
||||
use_ffn: true
|
||||
macaron_ffn: true
|
||||
merge_conv_kernel: 31
|
||||
|
||||
# decoder related
|
||||
decoder: transformer
|
||||
decoder_conf:
|
||||
attention_heads: 4
|
||||
linear_units: 2048
|
||||
num_blocks: 6
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
self_attention_dropout_rate: 0.
|
||||
src_attention_dropout_rate: 0.
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
# hybrid CTC/attention
|
||||
model_conf:
|
||||
ctc_weight: 0.3
|
||||
lsm_weight: 0.1 # label smoothing option
|
||||
length_normalized_loss: false
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 180
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- acc
|
||||
- max
|
||||
keep_nbest_models: 10
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.001
|
||||
weight_decay: 0.000001
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 35000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 27
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_ratio_range:
|
||||
- 0.
|
||||
- 0.05
|
||||
num_time_mask: 10
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 10000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
normalize: None
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,105 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (author: Daniel Povey)
|
||||
# 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
remove_archive=false
|
||||
|
||||
if [ "$1" == --remove-archive ]; then
|
||||
remove_archive=true
|
||||
shift
|
||||
fi
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
|
||||
echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
|
||||
echo "With --remove-archive it will remove the archive after successfully un-tarring it."
|
||||
echo "<corpus-part> can be one of: data_aishell, resource_aishell."
|
||||
fi
|
||||
|
||||
data=$1
|
||||
url=$2
|
||||
part=$3
|
||||
|
||||
if [ ! -d "$data" ]; then
|
||||
echo "$0: no such directory $data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
part_ok=false
|
||||
list="data_aishell resource_aishell"
|
||||
for x in $list; do
|
||||
if [ "$part" == $x ]; then part_ok=true; fi
|
||||
done
|
||||
if ! $part_ok; then
|
||||
echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -z "$url" ]; then
|
||||
echo "$0: empty URL base."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -f $data/$part/.complete ]; then
|
||||
echo "$0: data part $part was already successfully extracted, nothing to do."
|
||||
exit 0;
|
||||
fi
|
||||
|
||||
# sizes of the archive files in bytes.
|
||||
sizes="15582913665 1246920"
|
||||
|
||||
if [ -f $data/$part.tgz ]; then
|
||||
size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
|
||||
size_ok=false
|
||||
for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
|
||||
if ! $size_ok; then
|
||||
echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
|
||||
echo "does not equal the size of one of the archives."
|
||||
rm $data/$part.tgz
|
||||
else
|
||||
echo "$data/$part.tgz exists and appears to be complete."
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ ! -f $data/$part.tgz ]; then
|
||||
if ! command -v wget >/dev/null; then
|
||||
echo "$0: wget is not installed."
|
||||
exit 1;
|
||||
fi
|
||||
full_url=$url/$part.tgz
|
||||
echo "$0: downloading data from $full_url. This may take some time, please be patient."
|
||||
|
||||
cd $data || exit 1
|
||||
if ! wget --no-check-certificate $full_url; then
|
||||
echo "$0: error executing wget $full_url"
|
||||
exit 1;
|
||||
fi
|
||||
fi
|
||||
|
||||
cd $data || exit 1
|
||||
|
||||
if ! tar -xvzf $part.tgz; then
|
||||
echo "$0: error un-tarring archive $data/$part.tgz"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
touch $data/$part/.complete
|
||||
|
||||
if [ $part == "data_aishell" ]; then
|
||||
cd $data/$part/wav || exit 1
|
||||
for wav in ./*.tar.gz; do
|
||||
echo "Extracting wav from $wav"
|
||||
tar -zxf $wav && rm $wav
|
||||
done
|
||||
fi
|
||||
|
||||
echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
|
||||
|
||||
if $remove_archive; then
|
||||
echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
|
||||
rm $data/$part.tgz
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,225 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1,2,3"
|
||||
gpu_num=4
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=5
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_asr_e_branchformer.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_asr_transformer.yaml
|
||||
inference_asr_model=valid.acc.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/$token_type/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type $token_type \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode asr \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
|
||||
# Prepare files for ModelScope fine-tuning and inference
|
||||
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
|
||||
echo "stage 6: ModelScope Preparation"
|
||||
cp ${feats_dir}/data/${train_set}/cmvn/am.mvn ${exp_dir}/exp/${model_dir}/am.mvn
|
||||
vocab_size=$(cat ${token_list} | wc -l)
|
||||
python utils/gen_modelscope_configuration.py \
|
||||
--am_model_name $inference_asr_model \
|
||||
--mode asr \
|
||||
--model_name conformer \
|
||||
--dataset aishell \
|
||||
--output_dir $exp_dir/exp/$model_dir \
|
||||
--vocab_size $vocab_size \
|
||||
--tag $tag
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../transformer/utils
|
||||
@@ -1,24 +0,0 @@
|
||||
# Paraformer
|
||||
pretrained model in [ModelScope](https://www.modelscope.cn/home)：[speech_paraformer_asr_nat-aishell1-pytorch](https://www.modelscope.cn/models/damo/speech_paraformer_asr_nat-aishell1-pytorch/summary)
|
||||
|
||||
## Training Config
|
||||
- Feature info: using 80 dims fbank, global cmvn, speed perturb(0.9, 1.0, 1.1), specaugment
|
||||
- Train info: lr 5e-4, batch_size 25000, 2 gpu(Tesla V100), acc_grad 1, 50 epochs
|
||||
- Train config: conf/train_asr_paraformer_conformer_12e_6d_2048_256.yaml
|
||||
- LM config: LM was not used
|
||||
|
||||
## Results (CER)
|
||||
|
||||
- Decode config: conf/decode_asr_transformer_noctc_1best.yaml (ctc weight:0.0)
|
||||
|
||||
| testset | CER(%) |
|
||||
|:-----------:|:-------:|
|
||||
| dev | 4.66 |
|
||||
| test | 5.11 |
|
||||
|
||||
- Decode config: conf/decode_asr_transformer.yaml (ctc weight:0.5)
|
||||
|
||||
| testset | CER(%) |
|
||||
|:-----------:|:-------:|
|
||||
| dev | 4.52 |
|
||||
| test | 4.94 |
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 10
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.5
|
||||
lm_weight: 0.7
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 1
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.0
|
||||
lm_weight: 0.15
|
||||
@@ -1,109 +0,0 @@
|
||||
# network architecture
|
||||
# encoder related
|
||||
encoder: conformer
|
||||
encoder_conf:
|
||||
output_size: 256 # dimension of attention
|
||||
attention_heads: 4
|
||||
linear_units: 2048 # the number of units of position-wise feed forward
|
||||
num_blocks: 12 # the number of encoder blocks
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
attention_dropout_rate: 0.0
|
||||
input_layer: conv2d # encoder architecture type
|
||||
normalize_before: true
|
||||
pos_enc_layer_type: rel_pos
|
||||
selfattention_layer_type: rel_selfattn
|
||||
activation_type: swish
|
||||
macaron_style: true
|
||||
use_cnn_module: true
|
||||
cnn_module_kernel: 15
|
||||
|
||||
# decoder related
|
||||
decoder: paraformer_decoder_san
|
||||
decoder_conf:
|
||||
attention_heads: 4
|
||||
linear_units: 2048
|
||||
num_blocks: 6
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
self_attention_dropout_rate: 0.0
|
||||
src_attention_dropout_rate: 0.0
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
model: paraformer
|
||||
model_conf:
|
||||
ctc_weight: 0.3
|
||||
lsm_weight: 0.1
|
||||
length_normalized_loss: false
|
||||
predictor_weight: 1.0
|
||||
sampling_ratio: 0.4
|
||||
use_1st_decoder_loss: true
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 150
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- acc
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- acc
|
||||
- max
|
||||
keep_nbest_models: 10
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.0005
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 30000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 30
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_range:
|
||||
- 0
|
||||
- 40
|
||||
num_time_mask: 2
|
||||
|
||||
predictor: cif_predictor
|
||||
predictor_conf:
|
||||
idim: 256
|
||||
threshold: 1.0
|
||||
l_order: 1
|
||||
r_order: 1
|
||||
tail_threshold: 0.45
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 25000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
normalize: None
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,105 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (author: Daniel Povey)
|
||||
# 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
remove_archive=false
|
||||
|
||||
if [ "$1" == --remove-archive ]; then
|
||||
remove_archive=true
|
||||
shift
|
||||
fi
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
|
||||
echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
|
||||
echo "With --remove-archive it will remove the archive after successfully un-tarring it."
|
||||
echo "<corpus-part> can be one of: data_aishell, resource_aishell."
|
||||
fi
|
||||
|
||||
data=$1
|
||||
url=$2
|
||||
part=$3
|
||||
|
||||
if [ ! -d "$data" ]; then
|
||||
echo "$0: no such directory $data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
part_ok=false
|
||||
list="data_aishell resource_aishell"
|
||||
for x in $list; do
|
||||
if [ "$part" == $x ]; then part_ok=true; fi
|
||||
done
|
||||
if ! $part_ok; then
|
||||
echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -z "$url" ]; then
|
||||
echo "$0: empty URL base."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -f $data/$part/.complete ]; then
|
||||
echo "$0: data part $part was already successfully extracted, nothing to do."
|
||||
exit 0;
|
||||
fi
|
||||
|
||||
# sizes of the archive files in bytes.
|
||||
sizes="15582913665 1246920"
|
||||
|
||||
if [ -f $data/$part.tgz ]; then
|
||||
size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
|
||||
size_ok=false
|
||||
for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
|
||||
if ! $size_ok; then
|
||||
echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
|
||||
echo "does not equal the size of one of the archives."
|
||||
rm $data/$part.tgz
|
||||
else
|
||||
echo "$data/$part.tgz exists and appears to be complete."
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ ! -f $data/$part.tgz ]; then
|
||||
if ! command -v wget >/dev/null; then
|
||||
echo "$0: wget is not installed."
|
||||
exit 1;
|
||||
fi
|
||||
full_url=$url/$part.tgz
|
||||
echo "$0: downloading data from $full_url. This may take some time, please be patient."
|
||||
|
||||
cd $data || exit 1
|
||||
if ! wget --no-check-certificate $full_url; then
|
||||
echo "$0: error executing wget $full_url"
|
||||
exit 1;
|
||||
fi
|
||||
fi
|
||||
|
||||
cd $data || exit 1
|
||||
|
||||
if ! tar -xvzf $part.tgz; then
|
||||
echo "$0: error un-tarring archive $data/$part.tgz"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
touch $data/$part/.complete
|
||||
|
||||
if [ $part == "data_aishell" ]; then
|
||||
cd $data/$part/wav || exit 1
|
||||
for wav in ./*.tar.gz; do
|
||||
echo "Extracting wav from $wav"
|
||||
tar -zxf $wav && rm $wav
|
||||
done
|
||||
fi
|
||||
|
||||
echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
|
||||
|
||||
if $remove_archive; then
|
||||
echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
|
||||
rm $data/$part.tgz
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,227 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1"
|
||||
gpu_num=2
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=1
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_asr_paraformer_conformer_12e_6d_2048_256.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_asr_transformer_noctc_1best.yaml
|
||||
inference_asr_model=valid.acc.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/$token_type/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type $token_type \
|
||||
--token_list $token_list \
|
||||
--dataset_type small \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode paraformer \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
|
||||
# Prepare files for ModelScope fine-tuning and inference
|
||||
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
|
||||
echo "stage 6: ModelScope Preparation"
|
||||
cp ${feats_dir}/data/${train_set}/cmvn/am.mvn ${exp_dir}/exp/${model_dir}/am.mvn
|
||||
vocab_size=$(cat ${token_list} | wc -l)
|
||||
python utils/gen_modelscope_configuration.py \
|
||||
--am_model_name $inference_asr_model \
|
||||
--mode paraformer \
|
||||
--model_name paraformer \
|
||||
--dataset aishell \
|
||||
--output_dir $exp_dir/exp/$model_dir \
|
||||
--vocab_size $vocab_size \
|
||||
--nat _nat \
|
||||
--tag $tag
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../transformer/utils/
|
||||
@@ -1,18 +0,0 @@
|
||||
# ParaformerBert + specaug + speed perturbation + specaugmentation
|
||||
## Environments
|
||||
- date: `Mon Nov 21 13:25:30 CST 2022`
|
||||
- python version: `3.7.12`
|
||||
- FunASR version: `0.1.0`
|
||||
- pytorch version: `pytorch 1.7.0`
|
||||
|
||||
## Config files
|
||||
- train config: conf/train_asr_paraformerbert_conformer_12e_6d_2048_256.yaml
|
||||
- model size: 46M
|
||||
- lm config: LM was not used
|
||||
- decode config: conf/decode_asr_transformer_noctc_1best.yaml (CTC was not used)
|
||||
|
||||
## Results (CER)
|
||||
| testset | CER(%) |
|
||||
|:-----------:|:-------:|
|
||||
| dev | 4.30 |
|
||||
| test | 4.80 |
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 10
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.5
|
||||
lm_weight: 0.7
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 1
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.0
|
||||
lm_weight: 0.15
|
||||
-112
@@ -1,112 +0,0 @@
|
||||
# network architecture
|
||||
# encoder related
|
||||
encoder: conformer
|
||||
encoder_conf:
|
||||
output_size: 256 # dimension of attention
|
||||
attention_heads: 4
|
||||
linear_units: 2048 # the number of units of position-wise feed forward
|
||||
num_blocks: 12 # the number of encoder blocks
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
attention_dropout_rate: 0.0
|
||||
input_layer: conv2d # encoder architecture type
|
||||
normalize_before: true
|
||||
pos_enc_layer_type: rel_pos
|
||||
selfattention_layer_type: rel_selfattn
|
||||
activation_type: swish
|
||||
macaron_style: true
|
||||
use_cnn_module: true
|
||||
cnn_module_kernel: 15
|
||||
|
||||
# decoder related
|
||||
decoder: paraformer_decoder_san
|
||||
decoder_conf:
|
||||
attention_heads: 4
|
||||
linear_units: 2048
|
||||
num_blocks: 6
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
self_attention_dropout_rate: 0.0
|
||||
src_attention_dropout_rate: 0.0
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
# hybrid CTC/attention
|
||||
model: paraformer_bert
|
||||
model_conf:
|
||||
ctc_weight: 0.3
|
||||
lsm_weight: 0.1 # label smoothing option
|
||||
length_normalized_loss: false
|
||||
predictor_weight: 1.0
|
||||
sampling_ratio: 0.4
|
||||
embeds_id: 3
|
||||
embed_dims: 768
|
||||
embeds_loss_weight: 2.0
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 150
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- acc
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- acc
|
||||
- max
|
||||
keep_nbest_models: 10
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.0005
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 30000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 30
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_range:
|
||||
- 0
|
||||
- 40
|
||||
num_time_mask: 2
|
||||
|
||||
predictor: cif_predictor
|
||||
predictor_conf:
|
||||
idim: 256
|
||||
threshold: 1.0
|
||||
l_order: 1
|
||||
r_order: 1
|
||||
tail_threshold: 0.45
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text,embed
|
||||
data_types: sound,text,kaldi_ark
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 25000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
normalize: None
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,105 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (author: Daniel Povey)
|
||||
# 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
remove_archive=false
|
||||
|
||||
if [ "$1" == --remove-archive ]; then
|
||||
remove_archive=true
|
||||
shift
|
||||
fi
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
|
||||
echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
|
||||
echo "With --remove-archive it will remove the archive after successfully un-tarring it."
|
||||
echo "<corpus-part> can be one of: data_aishell, resource_aishell."
|
||||
fi
|
||||
|
||||
data=$1
|
||||
url=$2
|
||||
part=$3
|
||||
|
||||
if [ ! -d "$data" ]; then
|
||||
echo "$0: no such directory $data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
part_ok=false
|
||||
list="data_aishell resource_aishell"
|
||||
for x in $list; do
|
||||
if [ "$part" == $x ]; then part_ok=true; fi
|
||||
done
|
||||
if ! $part_ok; then
|
||||
echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -z "$url" ]; then
|
||||
echo "$0: empty URL base."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -f $data/$part/.complete ]; then
|
||||
echo "$0: data part $part was already successfully extracted, nothing to do."
|
||||
exit 0;
|
||||
fi
|
||||
|
||||
# sizes of the archive files in bytes.
|
||||
sizes="15582913665 1246920"
|
||||
|
||||
if [ -f $data/$part.tgz ]; then
|
||||
size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
|
||||
size_ok=false
|
||||
for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
|
||||
if ! $size_ok; then
|
||||
echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
|
||||
echo "does not equal the size of one of the archives."
|
||||
rm $data/$part.tgz
|
||||
else
|
||||
echo "$data/$part.tgz exists and appears to be complete."
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ ! -f $data/$part.tgz ]; then
|
||||
if ! command -v wget >/dev/null; then
|
||||
echo "$0: wget is not installed."
|
||||
exit 1;
|
||||
fi
|
||||
full_url=$url/$part.tgz
|
||||
echo "$0: downloading data from $full_url. This may take some time, please be patient."
|
||||
|
||||
cd $data || exit 1
|
||||
if ! wget --no-check-certificate $full_url; then
|
||||
echo "$0: error executing wget $full_url"
|
||||
exit 1;
|
||||
fi
|
||||
fi
|
||||
|
||||
cd $data || exit 1
|
||||
|
||||
if ! tar -xvzf $part.tgz; then
|
||||
echo "$0: error un-tarring archive $data/$part.tgz"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
touch $data/$part/.complete
|
||||
|
||||
if [ $part == "data_aishell" ]; then
|
||||
cd $data/$part/wav || exit 1
|
||||
for wav in ./*.tar.gz; do
|
||||
echo "Extracting wav from $wav"
|
||||
tar -zxf $wav && rm $wav
|
||||
done
|
||||
fi
|
||||
|
||||
echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
|
||||
|
||||
if $remove_archive; then
|
||||
echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
|
||||
rm $data/$part.tgz
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
@@ -1,62 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
stage=1
|
||||
stop_stage=3
|
||||
|
||||
bert_model_name="bert-base-chinese"
|
||||
raw_dataset_path="../DATA"
|
||||
nj=64
|
||||
model_path=${bert_model_name}
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
for data_set in train dev test;do
|
||||
scp=$raw_dataset_path/data/${data_set}/text
|
||||
local_scp_dir_raw=${raw_dataset_path}/data/embeds/${data_set}
|
||||
local_scp_dir=$local_scp_dir_raw/split$nj
|
||||
local_records_dir=$local_scp_dir_raw/ark
|
||||
|
||||
mkdir -p $local_records_dir
|
||||
mkdir -p $local_scp_dir
|
||||
|
||||
split_scps=""
|
||||
for JOB in $(seq ${nj}); do
|
||||
split_scps="$split_scps $local_scp_dir/data.$JOB.text"
|
||||
done
|
||||
|
||||
utils/split_scp.pl $scp ${split_scps}
|
||||
|
||||
|
||||
for num in {0..7};do
|
||||
tmp=`expr $num \* 4`
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
for idx in {1..4}; do
|
||||
JOB=`expr $tmp + $idx`
|
||||
echo "proces jobid=$JOB"
|
||||
{
|
||||
beg=0
|
||||
gpu=`expr $beg + $idx`
|
||||
echo ${local_scp_dir}/log.${JOB}
|
||||
python utils/extract_embeds.py $local_scp_dir/data.$JOB.text ${local_records_dir}/embeds.${JOB}.ark ${local_records_dir}/embeds.${JOB}.scp ${local_records_dir}/embeds.${JOB}.shape ${gpu} ${model_path} &> ${local_scp_dir}/log.${JOB}
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
done
|
||||
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
for JOB in $(seq ${nj}); do
|
||||
cat ${local_records_dir}/embeds.${JOB}.scp || exit 1;
|
||||
done > ${local_scp_dir_raw}/embeds.scp
|
||||
|
||||
for JOB in $(seq ${nj}); do
|
||||
cat ${local_records_dir}/embeds.${JOB}.shape || exit 1;
|
||||
done > ${local_scp_dir_raw}/embeds.shape
|
||||
fi
|
||||
|
||||
cp ${local_scp_dir_raw}/embeds.scp ${raw_dataset_path}/data/${data_set}/embeds.scp
|
||||
done
|
||||
|
||||
echo "embeds is in: ${local_scp_dir_raw}"
|
||||
echo "success"
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,237 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1"
|
||||
gpu_num=2
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=1
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary, for large data
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
skip_extract_embed=false
|
||||
bert_model_name="bert-base-chinese"
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_asr_paraformerbert_conformer_12e_6d_2048_256.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_asr_transformer_noctc_1best.yaml
|
||||
inference_asr_model=valid.acc.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/$token_type/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
if ! "${skip_extract_embed}"; then
|
||||
echo "extract embeddings..."
|
||||
local/extract_embeds.sh \
|
||||
--bert_model_name ${bert_model_name} \
|
||||
--raw_dataset_path ${feats_dir} \
|
||||
--nj $nj
|
||||
fi
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type $token_type \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text,embeds.scp" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode paraformer \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
|
||||
# Prepare files for ModelScope fine-tuning and inference
|
||||
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
|
||||
echo "stage 6: ModelScope Preparation"
|
||||
cp ${feats_dir}/data/${train_set}/cmvn/am.mvn ${exp_dir}/exp/${model_dir}/am.mvn
|
||||
vocab_size=$(cat ${token_list} | wc -l)
|
||||
python utils/gen_modelscope_configuration.py \
|
||||
--am_model_name $inference_asr_model \
|
||||
--mode paraformer \
|
||||
--model_name paraformer_bert \
|
||||
--dataset aishell \
|
||||
--output_dir $exp_dir/exp/$model_dir \
|
||||
--vocab_size $vocab_size \
|
||||
--nat _nat \
|
||||
--tag $tag
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../transformer/utils/
|
||||
@@ -1,18 +0,0 @@
|
||||
|
||||
# Streaming RNN-T Result
|
||||
|
||||
## Training Config
|
||||
- 8 gpu(Tesla V100)
|
||||
- Feature info: using 80 dims fbank, global cmvn, speed perturb(0.9, 1.0, 1.1), specaugment
|
||||
- Train config: conf/train_conformer_rnnt_unified.yaml
|
||||
- chunk config: chunk size 16, 1 left chunk
|
||||
- LM config: LM was not used
|
||||
- Model size: 90M
|
||||
|
||||
## Results (CER)
|
||||
- Decode config: conf/decode_rnnt_conformer_streaming.yaml
|
||||
|
||||
| testset | CER(%) |
|
||||
|:-----------:|:-------:|
|
||||
| dev | 5.43 |
|
||||
| test | 6.04 |
|
||||
@@ -1,8 +0,0 @@
|
||||
# The conformer transducer decoding configuration from @jeon30c
|
||||
beam_size: 10
|
||||
simu_streaming: false
|
||||
streaming: true
|
||||
chunk_size: 16
|
||||
left_context: 16
|
||||
right_context: 0
|
||||
|
||||
@@ -1,5 +0,0 @@
|
||||
# The conformer transducer decoding configuration from @jeon30c
|
||||
beam_size: 10
|
||||
simu_streaming: true
|
||||
streaming: false
|
||||
chunk_size: 16
|
||||
@@ -1,99 +0,0 @@
|
||||
encoder: chunk_conformer
|
||||
encoder_conf:
|
||||
activation_type: swish
|
||||
positional_dropout_rate: 0.5
|
||||
time_reduction_factor: 2
|
||||
unified_model_training: true
|
||||
default_chunk_size: 16
|
||||
jitter_range: 4
|
||||
left_chunk_size: 1
|
||||
embed_vgg_like: false
|
||||
subsampling_factor: 4
|
||||
linear_units: 2048
|
||||
output_size: 512
|
||||
attention_heads: 8
|
||||
dropout_rate: 0.5
|
||||
positional_dropout_rate: 0.5
|
||||
attention_dropout_rate: 0.5
|
||||
cnn_module_kernel: 15
|
||||
num_blocks: 12
|
||||
|
||||
# decoder related
|
||||
rnnt_decoder: rnnt
|
||||
rnnt_decoder_conf:
|
||||
embed_size: 512
|
||||
hidden_size: 512
|
||||
embed_dropout_rate: 0.5
|
||||
dropout_rate: 0.5
|
||||
joint_network_conf:
|
||||
joint_space_size: 512
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
|
||||
# Auxiliary CTC
|
||||
model: rnnt_unified
|
||||
model_conf:
|
||||
auxiliary_ctc_weight: 0.0
|
||||
|
||||
# minibatch related
|
||||
use_amp: true
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 120
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- loss
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- cer_transducer_chunk
|
||||
- min
|
||||
keep_nbest_models: 10
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.001
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 25000
|
||||
|
||||
specaug: specaug
|
||||
specaug_conf:
|
||||
apply_time_warp: true
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 40
|
||||
num_freq_mask: 2
|
||||
apply_time_mask: true
|
||||
time_mask_width_range:
|
||||
- 0
|
||||
- 50
|
||||
num_time_mask: 5
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 16000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,5 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
@@ -1,210 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
|
||||
gpu_num=8
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=5
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_conformer_rnnt_unified.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_rnnt_conformer_streaming.yaml
|
||||
inference_asr_model=valid.cer_transducer_chunk.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} ${feats_dir}/data/${train_set}
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/char/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/char/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=./ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type char \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/cmvn.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/cmvn.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode rnnt \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
@@ -1 +0,0 @@
|
||||
../transformer/utils
|
||||
@@ -1,6 +0,0 @@
|
||||
beam_size: 10
|
||||
penalty: 0.0
|
||||
maxlenratio: 0.0
|
||||
minlenratio: 0.0
|
||||
ctc_weight: 0.5
|
||||
lm_weight: 0.7
|
||||
@@ -1,88 +0,0 @@
|
||||
# network architecture
|
||||
# encoder related
|
||||
encoder: transformer
|
||||
encoder_conf:
|
||||
output_size: 256 # dimension of attention
|
||||
attention_heads: 4
|
||||
linear_units: 2048 # the number of units of position-wise feed forward
|
||||
num_blocks: 12 # the number of encoder blocks
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
attention_dropout_rate: 0.0
|
||||
input_layer: conv2d # encoder architecture type
|
||||
normalize_before: true
|
||||
|
||||
# decoder related
|
||||
decoder: transformer
|
||||
decoder_conf:
|
||||
attention_heads: 4
|
||||
linear_units: 2048
|
||||
num_blocks: 6
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
self_attention_dropout_rate: 0.0
|
||||
src_attention_dropout_rate: 0.0
|
||||
|
||||
# frontend related
|
||||
frontend: wav_frontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 1
|
||||
lfr_n: 1
|
||||
|
||||
# hybrid CTC/attention
|
||||
model_conf:
|
||||
ctc_weight: 0.3
|
||||
lsm_weight: 0.1 # label smoothing option
|
||||
length_normalized_loss: false
|
||||
|
||||
# optimization related
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
patience: none
|
||||
max_epoch: 60
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- acc
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- acc
|
||||
- max
|
||||
keep_nbest_models: 10
|
||||
|
||||
# NoamLR is deprecated. Use WarmupLR.
|
||||
# The following is equivalent setting for NoamLR:
|
||||
#
|
||||
# optim: adam
|
||||
# optim_conf:
|
||||
# lr: 10.
|
||||
# scheduler: noamlr
|
||||
# scheduler_conf:
|
||||
# model_size: 256
|
||||
# warmup_steps: 25000
|
||||
#
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.002
|
||||
scheduler: warmuplr # pytorch v1.1.0+ required
|
||||
scheduler_conf:
|
||||
warmup_steps: 25000
|
||||
|
||||
dataset_conf:
|
||||
data_names: speech,text
|
||||
data_types: sound,text
|
||||
shuffle: True
|
||||
shuffle_conf:
|
||||
shuffle_size: 2048
|
||||
sort_size: 500
|
||||
batch_conf:
|
||||
batch_type: token
|
||||
batch_size: 25000
|
||||
num_workers: 8
|
||||
|
||||
log_interval: 50
|
||||
normalize: None
|
||||
@@ -1,66 +0,0 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Copyright 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
#. ./path.sh || exit 1;
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 <audio-path> <text-path> <output-path>"
|
||||
echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
aishell_audio_dir=$1
|
||||
aishell_text=$2/aishell_transcript_v0.8.txt
|
||||
output_dir=$3
|
||||
|
||||
train_dir=$output_dir/data/local/train
|
||||
dev_dir=$output_dir/data/local/dev
|
||||
test_dir=$output_dir/data/local/test
|
||||
tmp_dir=$output_dir/data/local/tmp
|
||||
|
||||
mkdir -p $train_dir
|
||||
mkdir -p $dev_dir
|
||||
mkdir -p $test_dir
|
||||
mkdir -p $tmp_dir
|
||||
|
||||
# data directory check
|
||||
if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then
|
||||
echo "Error: $0 requires two directory arguments"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
# find wav audio file for train, dev and test resp.
|
||||
find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist
|
||||
n=`cat $tmp_dir/wav.flist | wc -l`
|
||||
[ $n -ne 141925 ] && \
|
||||
echo Warning: expected 141925 data data files, found $n
|
||||
|
||||
grep -i "wav/train" $tmp_dir/wav.flist > $train_dir/wav.flist || exit 1;
|
||||
grep -i "wav/dev" $tmp_dir/wav.flist > $dev_dir/wav.flist || exit 1;
|
||||
grep -i "wav/test" $tmp_dir/wav.flist > $test_dir/wav.flist || exit 1;
|
||||
|
||||
rm -r $tmp_dir
|
||||
|
||||
# Transcriptions preparation
|
||||
for dir in $train_dir $dev_dir $test_dir; do
|
||||
echo Preparing $dir transcriptions
|
||||
sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list
|
||||
paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt
|
||||
awk '{print $1}' $dir/transcripts.txt > $dir/utt.list
|
||||
utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp
|
||||
sort -u $dir/transcripts.txt > $dir/text
|
||||
done
|
||||
|
||||
mkdir -p $output_dir/data/train $output_dir/data/dev $output_dir/data/test
|
||||
|
||||
for f in wav.scp text; do
|
||||
cp $train_dir/$f $output_dir/data/train/$f || exit 1;
|
||||
cp $dev_dir/$f $output_dir/data/dev/$f || exit 1;
|
||||
cp $test_dir/$f $output_dir/data/test/$f || exit 1;
|
||||
done
|
||||
|
||||
echo "$0: AISHELL data preparation succeeded"
|
||||
exit 0;
|
||||
@@ -1,105 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (author: Daniel Povey)
|
||||
# 2017 Xingyu Na
|
||||
# Apache 2.0
|
||||
|
||||
remove_archive=false
|
||||
|
||||
if [ "$1" == --remove-archive ]; then
|
||||
remove_archive=true
|
||||
shift
|
||||
fi
|
||||
|
||||
if [ $# -ne 3 ]; then
|
||||
echo "Usage: $0 [--remove-archive] <data-base> <url-base> <corpus-part>"
|
||||
echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/33 data_aishell"
|
||||
echo "With --remove-archive it will remove the archive after successfully un-tarring it."
|
||||
echo "<corpus-part> can be one of: data_aishell, resource_aishell."
|
||||
fi
|
||||
|
||||
data=$1
|
||||
url=$2
|
||||
part=$3
|
||||
|
||||
if [ ! -d "$data" ]; then
|
||||
echo "$0: no such directory $data"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
part_ok=false
|
||||
list="data_aishell resource_aishell"
|
||||
for x in $list; do
|
||||
if [ "$part" == $x ]; then part_ok=true; fi
|
||||
done
|
||||
if ! $part_ok; then
|
||||
echo "$0: expected <corpus-part> to be one of $list, but got '$part'"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -z "$url" ]; then
|
||||
echo "$0: empty URL base."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ -f $data/$part/.complete ]; then
|
||||
echo "$0: data part $part was already successfully extracted, nothing to do."
|
||||
exit 0;
|
||||
fi
|
||||
|
||||
# sizes of the archive files in bytes.
|
||||
sizes="15582913665 1246920"
|
||||
|
||||
if [ -f $data/$part.tgz ]; then
|
||||
size=$(/bin/ls -l $data/$part.tgz | awk '{print $5}')
|
||||
size_ok=false
|
||||
for s in $sizes; do if [ $s == $size ]; then size_ok=true; fi; done
|
||||
if ! $size_ok; then
|
||||
echo "$0: removing existing file $data/$part.tgz because its size in bytes $size"
|
||||
echo "does not equal the size of one of the archives."
|
||||
rm $data/$part.tgz
|
||||
else
|
||||
echo "$data/$part.tgz exists and appears to be complete."
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ ! -f $data/$part.tgz ]; then
|
||||
if ! command -v wget >/dev/null; then
|
||||
echo "$0: wget is not installed."
|
||||
exit 1;
|
||||
fi
|
||||
full_url=$url/$part.tgz
|
||||
echo "$0: downloading data from $full_url. This may take some time, please be patient."
|
||||
|
||||
cd $data || exit 1
|
||||
if ! wget --no-check-certificate $full_url; then
|
||||
echo "$0: error executing wget $full_url"
|
||||
exit 1;
|
||||
fi
|
||||
fi
|
||||
|
||||
cd $data || exit 1
|
||||
|
||||
if ! tar -xvzf $part.tgz; then
|
||||
echo "$0: error un-tarring archive $data/$part.tgz"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
touch $data/$part/.complete
|
||||
|
||||
if [ $part == "data_aishell" ]; then
|
||||
cd $data/$part/wav || exit 1
|
||||
for wav in ./*.tar.gz; do
|
||||
echo "Extracting wav from $wav"
|
||||
tar -zxf $wav && rm $wav
|
||||
done
|
||||
fi
|
||||
|
||||
echo "$0: Successfully downloaded and un-tarred $data/$part.tgz"
|
||||
|
||||
if $remove_archive; then
|
||||
echo "$0: removing $data/$part.tgz file since --remove-archive option was supplied."
|
||||
rm $data/$part.tgz
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
@@ -1,7 +0,0 @@
|
||||
export FUNASR_DIR=$PWD/../../..
|
||||
|
||||
# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
|
||||
export PYTHONIOENCODING=UTF-8
|
||||
export PATH=$FUNASR_DIR/funasr/bin:$PATH
|
||||
|
||||
export OMP_NUM_THREADS=1
|
||||
@@ -1,225 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
|
||||
# machines configuration
|
||||
CUDA_VISIBLE_DEVICES="0,1"
|
||||
gpu_num=2
|
||||
count=1
|
||||
gpu_inference=true # Whether to perform gpu decoding, set false for cpu decoding
|
||||
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
|
||||
njob=5
|
||||
train_cmd=utils/run.pl
|
||||
infer_cmd=utils/run.pl
|
||||
|
||||
# general configuration
|
||||
feats_dir="../DATA" #feature output dictionary
|
||||
exp_dir="."
|
||||
lang=zh
|
||||
token_type=char
|
||||
type=sound
|
||||
scp=wav.scp
|
||||
speed_perturb="0.9 1.0 1.1"
|
||||
stage=0
|
||||
stop_stage=5
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
nj=64
|
||||
|
||||
# data
|
||||
raw_data=../raw_data
|
||||
data_url=www.openslr.org/resources/33
|
||||
|
||||
# exp tag
|
||||
tag="exp1"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# Set bash to 'debug' mode, it will exit on :
|
||||
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
|
||||
set -e
|
||||
set -u
|
||||
set -o pipefail
|
||||
|
||||
train_set=train
|
||||
valid_set=dev
|
||||
test_sets="dev test"
|
||||
|
||||
asr_config=conf/train_asr_transformer.yaml
|
||||
model_dir="baseline_$(basename "${asr_config}" .yaml)_${lang}_${token_type}_${tag}"
|
||||
|
||||
inference_config=conf/decode_asr_transformer.yaml
|
||||
inference_asr_model=valid.acc.ave_10best.pb
|
||||
|
||||
# you can set gpu num for decoding here
|
||||
gpuid_list=$CUDA_VISIBLE_DEVICES # set gpus for decoding, the same as training stage by default
|
||||
ngpu=$(echo $gpuid_list | awk -F "," '{print NF}')
|
||||
|
||||
if ${gpu_inference}; then
|
||||
inference_nj=$[${ngpu}*${njob}]
|
||||
_ngpu=1
|
||||
else
|
||||
inference_nj=$njob
|
||||
_ngpu=0
|
||||
fi
|
||||
|
||||
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
|
||||
echo "stage -1: Data Download"
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} data_aishell
|
||||
local/download_and_untar.sh ${raw_data} ${data_url} resource_aishell
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
echo "stage 0: Data preparation"
|
||||
# Data preparation
|
||||
local/aishell_data_prep.sh ${raw_data}/data_aishell/wav ${raw_data}/data_aishell/transcript ${feats_dir}
|
||||
for x in train dev test; do
|
||||
cp ${feats_dir}/data/${x}/text ${feats_dir}/data/${x}/text.org
|
||||
paste -d " " <(cut -f 1 -d" " ${feats_dir}/data/${x}/text.org) <(cut -f 2- -d" " ${feats_dir}/data/${x}/text.org | tr -d " ") \
|
||||
> ${feats_dir}/data/${x}/text
|
||||
utils/text2token.py -n 1 -s 1 ${feats_dir}/data/${x}/text > ${feats_dir}/data/${x}/text.org
|
||||
mv ${feats_dir}/data/${x}/text.org ${feats_dir}/data/${x}/text
|
||||
done
|
||||
fi
|
||||
|
||||
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
|
||||
echo "stage 1: Feature and CMVN Generation"
|
||||
utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 1.0
|
||||
fi
|
||||
|
||||
token_list=${feats_dir}/data/${lang}_token_list/$token_type/tokens.txt
|
||||
echo "dictionary: ${token_list}"
|
||||
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
|
||||
echo "stage 2: Dictionary Preparation"
|
||||
mkdir -p ${feats_dir}/data/${lang}_token_list/$token_type/
|
||||
|
||||
echo "make a dictionary"
|
||||
echo "<blank>" > ${token_list}
|
||||
echo "<s>" >> ${token_list}
|
||||
echo "</s>" >> ${token_list}
|
||||
utils/text2token.py -s 1 -n 1 --space "" ${feats_dir}/data/$train_set/text | cut -f 2- -d" " | tr " " "\n" \
|
||||
| sort | uniq | grep -a -v -e '^\s*$' | awk '{print $0}' >> ${token_list}
|
||||
echo "<unk>" >> ${token_list}
|
||||
fi
|
||||
|
||||
# LM Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
|
||||
echo "stage 3: LM Training"
|
||||
fi
|
||||
|
||||
# ASR Training Stage
|
||||
world_size=$gpu_num # run on one machine
|
||||
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
|
||||
echo "stage 4: ASR Training"
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}
|
||||
mkdir -p ${exp_dir}/exp/${model_dir}/log
|
||||
INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
|
||||
if [ -f $INIT_FILE ];then
|
||||
rm -f $INIT_FILE
|
||||
fi
|
||||
init_method=file://$(readlink -f $INIT_FILE)
|
||||
echo "$0: init method is $init_method"
|
||||
for ((i = 0; i < $gpu_num; ++i)); do
|
||||
{
|
||||
rank=$i
|
||||
local_rank=$i
|
||||
gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
|
||||
train.py \
|
||||
--task_name asr \
|
||||
--gpu_id $gpu_id \
|
||||
--use_preprocessor true \
|
||||
--token_type $token_type \
|
||||
--token_list $token_list \
|
||||
--data_dir ${feats_dir}/data \
|
||||
--train_set ${train_set} \
|
||||
--valid_set ${valid_set} \
|
||||
--data_file_names "wav.scp,text" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--speed_perturb ${speed_perturb} \
|
||||
--resume true \
|
||||
--output_dir ${exp_dir}/exp/${model_dir} \
|
||||
--config $asr_config \
|
||||
--ngpu $gpu_num \
|
||||
--num_worker_count $count \
|
||||
--dist_init_method $init_method \
|
||||
--dist_world_size $world_size \
|
||||
--dist_rank $rank \
|
||||
--local_rank $local_rank 1> ${exp_dir}/exp/${model_dir}/log/train.log.$i 2>&1
|
||||
} &
|
||||
done
|
||||
wait
|
||||
fi
|
||||
|
||||
# Testing Stage
|
||||
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
|
||||
echo "stage 5: Inference"
|
||||
for dset in ${test_sets}; do
|
||||
asr_exp=${exp_dir}/exp/${model_dir}
|
||||
inference_tag="$(basename "${inference_config}" .yaml)"
|
||||
_dir="${asr_exp}/${inference_tag}/${inference_asr_model}/${dset}"
|
||||
_logdir="${_dir}/logdir"
|
||||
if [ -d ${_dir} ]; then
|
||||
echo "${_dir} is already exists. if you want to decode again, please delete this dir first."
|
||||
exit 0
|
||||
fi
|
||||
mkdir -p "${_logdir}"
|
||||
_data="${feats_dir}/data/${dset}"
|
||||
key_file=${_data}/${scp}
|
||||
num_scp_file="$(<${key_file} wc -l)"
|
||||
_nj=$([ $inference_nj -le $num_scp_file ] && echo "$inference_nj" || echo "$num_scp_file")
|
||||
split_scps=
|
||||
for n in $(seq "${_nj}"); do
|
||||
split_scps+=" ${_logdir}/keys.${n}.scp"
|
||||
done
|
||||
# shellcheck disable=SC2086
|
||||
utils/split_scp.pl "${key_file}" ${split_scps}
|
||||
_opts=
|
||||
if [ -n "${inference_config}" ]; then
|
||||
_opts+="--config ${inference_config} "
|
||||
fi
|
||||
${infer_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
|
||||
python -m funasr.bin.asr_inference_launch \
|
||||
--batch_size 1 \
|
||||
--ngpu "${_ngpu}" \
|
||||
--njob ${njob} \
|
||||
--gpuid_list ${gpuid_list} \
|
||||
--data_path_and_name_and_type "${_data}/${scp},speech,${type}" \
|
||||
--cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
|
||||
--key_file "${_logdir}"/keys.JOB.scp \
|
||||
--asr_train_config "${asr_exp}"/config.yaml \
|
||||
--asr_model_file "${asr_exp}"/"${inference_asr_model}" \
|
||||
--output_dir "${_logdir}"/output.JOB \
|
||||
--mode asr \
|
||||
${_opts}
|
||||
|
||||
for f in token token_int score text; do
|
||||
if [ -f "${_logdir}/output.1/1best_recog/${f}" ]; then
|
||||
for i in $(seq "${_nj}"); do
|
||||
cat "${_logdir}/output.${i}/1best_recog/${f}"
|
||||
done | sort -k1 >"${_dir}/${f}"
|
||||
fi
|
||||
done
|
||||
python utils/proce_text.py ${_dir}/text ${_dir}/text.proc
|
||||
python utils/proce_text.py ${_data}/text ${_data}/text.proc
|
||||
python utils/compute_wer.py ${_data}/text.proc ${_dir}/text.proc ${_dir}/text.cer
|
||||
tail -n 3 ${_dir}/text.cer > ${_dir}/text.cer.txt
|
||||
cat ${_dir}/text.cer.txt
|
||||
done
|
||||
fi
|
||||
|
||||
# Prepare files for ModelScope fine-tuning and inference
|
||||
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
|
||||
echo "stage 6: ModelScope Preparation"
|
||||
cp ${feats_dir}/data/${train_set}/cmvn/am.mvn ${exp_dir}/exp/${model_dir}/am.mvn
|
||||
vocab_size=$(cat ${token_list} | wc -l)
|
||||
python utils/gen_modelscope_configuration.py \
|
||||
--am_model_name $inference_asr_model \
|
||||
--mode asr \
|
||||
--model_name transformer \
|
||||
--dataset aishell \
|
||||
--output_dir $exp_dir/exp/$model_dir \
|
||||
--vocab_size $vocab_size \
|
||||
--tag $tag
|
||||
fi
|
||||
@@ -1,79 +0,0 @@
|
||||
from kaldiio import ReadHelper
|
||||
from kaldiio import WriteHelper
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import math
|
||||
import numpy as np
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="apply cmvn",
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
|
||||
)
|
||||
parser.add_argument(
|
||||
"--ark-file",
|
||||
"-a",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="fbank ark file",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--cmvn-file",
|
||||
"-c",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="cmvn file",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--ark-index",
|
||||
"-i",
|
||||
default=1,
|
||||
required=True,
|
||||
type=int,
|
||||
help="ark index",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--output-dir",
|
||||
"-o",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="output dir",
|
||||
)
|
||||
return parser
|
||||
|
||||
|
||||
def main():
|
||||
parser = get_parser()
|
||||
args = parser.parse_args()
|
||||
|
||||
ark_file = args.output_dir + "/feats." + str(args.ark_index) + ".ark"
|
||||
scp_file = args.output_dir + "/feats." + str(args.ark_index) + ".scp"
|
||||
ark_writer = WriteHelper('ark,scp:{},{}'.format(ark_file, scp_file))
|
||||
|
||||
with open(args.cmvn_file) as f:
|
||||
cmvn_stats = json.load(f)
|
||||
|
||||
means = cmvn_stats['mean_stats']
|
||||
vars = cmvn_stats['var_stats']
|
||||
total_frames = cmvn_stats['total_frames']
|
||||
|
||||
for i in range(len(means)):
|
||||
means[i] /= total_frames
|
||||
vars[i] = vars[i] / total_frames - means[i] * means[i]
|
||||
if vars[i] < 1.0e-20:
|
||||
vars[i] = 1.0e-20
|
||||
vars[i] = 1.0 / math.sqrt(vars[i])
|
||||
|
||||
with ReadHelper('ark:{}'.format(args.ark_file)) as ark_reader:
|
||||
for key, mat in ark_reader:
|
||||
mat = (mat - means) * vars
|
||||
ark_writer(key, mat)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
@@ -1,36 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
# Begin configuration section.
|
||||
nj=32
|
||||
cmd=./utils/run.pl
|
||||
|
||||
echo "$0 $@"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
fbankdir=$1
|
||||
cmvn_file=$2
|
||||
logdir=$3
|
||||
output_dir=$4
|
||||
|
||||
dump_dir=${output_dir}/ark; mkdir -p ${dump_dir}
|
||||
mkdir -p ${logdir}
|
||||
|
||||
#$cmd JOB=1:$nj $logdir/apply_cmvn.JOB.log \
|
||||
# python utils/apply_cmvn.py -a $fbankdir/ark/feats.JOB.ark \
|
||||
# -c $cmvn_file -i JOB -o ${dump_dir} \
|
||||
# || exit 1;
|
||||
|
||||
for JOB in `seq 1 $nj`;do
|
||||
{
|
||||
python utils/apply_cmvn.py -a $fbankdir/ark/feats.${JOB}.ark \
|
||||
-c $cmvn_file -i ${JOB} -o ${dump_dir} || exit 1;
|
||||
} &> $logdir/apply_cmvn.${JOB}.log &
|
||||
done
|
||||
|
||||
for n in $(seq $nj); do
|
||||
cat ${dump_dir}/feats.$n.scp || exit 1
|
||||
done > ${output_dir}/feats.scp || exit 1
|
||||
|
||||
echo "$0: Succeeded apply cmvn"
|
||||
@@ -1,143 +0,0 @@
|
||||
from kaldiio import ReadHelper, WriteHelper
|
||||
|
||||
import argparse
|
||||
import numpy as np
|
||||
|
||||
|
||||
def build_LFR_features(inputs, m=7, n=6):
|
||||
LFR_inputs = []
|
||||
T = inputs.shape[0]
|
||||
T_lfr = int(np.ceil(T / n))
|
||||
left_padding = np.tile(inputs[0], ((m - 1) // 2, 1))
|
||||
inputs = np.vstack((left_padding, inputs))
|
||||
T = T + (m - 1) // 2
|
||||
for i in range(T_lfr):
|
||||
if m <= T - i * n:
|
||||
LFR_inputs.append(np.hstack(inputs[i * n:i * n + m]))
|
||||
else:
|
||||
num_padding = m - (T - i * n)
|
||||
frame = np.hstack(inputs[i * n:])
|
||||
for _ in range(num_padding):
|
||||
frame = np.hstack((frame, inputs[-1]))
|
||||
LFR_inputs.append(frame)
|
||||
return np.vstack(LFR_inputs)
|
||||
|
||||
|
||||
def build_CMVN_features(inputs, mvn_file): # noqa
|
||||
with open(mvn_file, 'r', encoding='utf-8') as f:
|
||||
lines = f.readlines()
|
||||
|
||||
add_shift_list = []
|
||||
rescale_list = []
|
||||
for i in range(len(lines)):
|
||||
line_item = lines[i].split()
|
||||
if line_item[0] == '<AddShift>':
|
||||
line_item = lines[i + 1].split()
|
||||
if line_item[0] == '<LearnRateCoef>':
|
||||
add_shift_line = line_item[3:(len(line_item) - 1)]
|
||||
add_shift_list = list(add_shift_line)
|
||||
continue
|
||||
elif line_item[0] == '<Rescale>':
|
||||
line_item = lines[i + 1].split()
|
||||
if line_item[0] == '<LearnRateCoef>':
|
||||
rescale_line = line_item[3:(len(line_item) - 1)]
|
||||
rescale_list = list(rescale_line)
|
||||
continue
|
||||
|
||||
for j in range(inputs.shape[0]):
|
||||
for k in range(inputs.shape[1]):
|
||||
add_shift_value = add_shift_list[k]
|
||||
rescale_value = rescale_list[k]
|
||||
inputs[j, k] = float(inputs[j, k]) + float(add_shift_value)
|
||||
inputs[j, k] = float(inputs[j, k]) * float(rescale_value)
|
||||
|
||||
return inputs
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="apply low_frame_rate and cmvn",
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
|
||||
)
|
||||
parser.add_argument(
|
||||
"--ark-file",
|
||||
"-a",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="fbank ark file",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--lfr",
|
||||
"-f",
|
||||
default=True,
|
||||
type=str,
|
||||
help="low frame rate",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--lfr-m",
|
||||
"-m",
|
||||
default=7,
|
||||
type=int,
|
||||
help="number of frames to stack",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--lfr-n",
|
||||
"-n",
|
||||
default=6,
|
||||
type=int,
|
||||
help="number of frames to skip",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--cmvn-file",
|
||||
"-c",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="global cmvn file",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--ark-index",
|
||||
"-i",
|
||||
default=1,
|
||||
required=True,
|
||||
type=int,
|
||||
help="ark index",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--output-dir",
|
||||
"-o",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="output dir",
|
||||
)
|
||||
return parser
|
||||
|
||||
|
||||
def main():
|
||||
parser = get_parser()
|
||||
args = parser.parse_args()
|
||||
|
||||
dump_ark_file = args.output_dir + "/feats." + str(args.ark_index) + ".ark"
|
||||
dump_scp_file = args.output_dir + "/feats." + str(args.ark_index) + ".scp"
|
||||
shape_file = args.output_dir + "/len." + str(args.ark_index)
|
||||
ark_writer = WriteHelper('ark,scp:{},{}'.format(dump_ark_file, dump_scp_file))
|
||||
|
||||
shape_writer = open(shape_file, 'w')
|
||||
with ReadHelper('ark:{}'.format(args.ark_file)) as ark_reader:
|
||||
for key, mat in ark_reader:
|
||||
if args.lfr:
|
||||
lfr = build_LFR_features(mat, args.lfr_m, args.lfr_n)
|
||||
else:
|
||||
lfr = mat
|
||||
cmvn = build_CMVN_features(lfr, args.cmvn_file)
|
||||
dims = cmvn.shape[1]
|
||||
lens = cmvn.shape[0]
|
||||
shape_writer.write(key + " " + str(lens) + "," + str(dims) + '\n')
|
||||
ark_writer(key, cmvn)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
|
||||
@@ -1,38 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
|
||||
# Begin configuration section.
|
||||
nj=32
|
||||
cmd=utils/run.pl
|
||||
|
||||
# feature configuration
|
||||
lfr=True
|
||||
lfr_m=7
|
||||
lfr_n=6
|
||||
|
||||
echo "$0 $@"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
fbankdir=$1
|
||||
cmvn_file=$2
|
||||
logdir=$3
|
||||
output_dir=$4
|
||||
|
||||
dump_dir=${output_dir}/ark; mkdir -p ${dump_dir}
|
||||
mkdir -p ${logdir}
|
||||
|
||||
$cmd JOB=1:$nj $logdir/apply_lfr_and_cmvn.JOB.log \
|
||||
python utils/apply_lfr_and_cmvn.py -a $fbankdir/ark/feats.JOB.ark \
|
||||
-f $lfr -m $lfr_m -n $lfr_n -c $cmvn_file -i JOB -o ${dump_dir} \
|
||||
|| exit 1;
|
||||
|
||||
for n in $(seq $nj); do
|
||||
cat ${dump_dir}/feats.$n.scp || exit 1
|
||||
done > ${output_dir}/feats.scp || exit 1
|
||||
|
||||
for n in $(seq $nj); do
|
||||
cat ${dump_dir}/len.$n || exit 1
|
||||
done > ${output_dir}/speech_shape || exit 1
|
||||
|
||||
echo "$0: Succeeded apply low frame rate and cmvn"
|
||||
@@ -1,51 +0,0 @@
|
||||
import argparse
|
||||
import json
|
||||
import numpy as np
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="cmvn converter",
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
|
||||
)
|
||||
parser.add_argument(
|
||||
"--cmvn_json",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="cmvn json file",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--am_mvn",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="am mvn file",
|
||||
)
|
||||
return parser
|
||||
|
||||
def main():
|
||||
parser = get_parser()
|
||||
args = parser.parse_args()
|
||||
|
||||
with open(args.cmvn_json, "r") as fin:
|
||||
cmvn_dict = json.load(fin)
|
||||
|
||||
mean_stats = np.array(cmvn_dict["mean_stats"])
|
||||
var_stats = np.array(cmvn_dict["var_stats"])
|
||||
total_frame = np.array(cmvn_dict["total_frames"])
|
||||
|
||||
mean = -1.0 * mean_stats / total_frame
|
||||
var = 1.0 / np.sqrt(var_stats / total_frame - mean * mean)
|
||||
dims = mean.shape[0]
|
||||
with open(args.am_mvn, 'w') as fout:
|
||||
fout.write("<Nnet>" + "\n" + "<Splice> " + str(dims) + " " + str(dims) + '\n' + "[ 0 ]" + "\n" + "<AddShift> " + str(dims) + " " + str(dims) + "\n")
|
||||
mean_str = str(list(mean)).replace(',', '').replace('[', '[ ').replace(']', ' ]')
|
||||
fout.write("<LearnRateCoef> 0 " + mean_str + '\n')
|
||||
fout.write("<Rescale> " + str(dims) + " " + str(dims) + '\n')
|
||||
var_str = str(list(var)).replace(',', '').replace('[', '[ ').replace(']', ' ]')
|
||||
fout.write("<LearnRateCoef> 0 " + var_str + '\n')
|
||||
fout.write("</Nnet>" + '\n')
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
@@ -1,72 +0,0 @@
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
|
||||
import numpy as np
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="combine cmvn file",
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
|
||||
)
|
||||
parser.add_argument(
|
||||
"--dim",
|
||||
default=80,
|
||||
type=int,
|
||||
help="feature dim",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--cmvn_dir",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="cmvn dir",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--nj",
|
||||
default=1,
|
||||
required=True,
|
||||
type=int,
|
||||
help="num of cmvn files",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--output_dir",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="output dir",
|
||||
)
|
||||
return parser
|
||||
|
||||
|
||||
def main():
|
||||
parser = get_parser()
|
||||
args = parser.parse_args()
|
||||
|
||||
total_means = np.zeros(args.dim)
|
||||
total_vars = np.zeros(args.dim)
|
||||
total_frames = 0
|
||||
|
||||
cmvn_file = os.path.join(args.output_dir, "cmvn.json")
|
||||
|
||||
for i in range(1, args.nj + 1):
|
||||
with open(os.path.join(args.cmvn_dir, "cmvn.{}.json".format(str(i)))) as fin:
|
||||
cmvn_stats = json.load(fin)
|
||||
|
||||
total_means += np.array(cmvn_stats["mean_stats"])
|
||||
total_vars += np.array(cmvn_stats["var_stats"])
|
||||
total_frames += cmvn_stats["total_frames"]
|
||||
|
||||
cmvn_info = {
|
||||
'mean_stats': list(total_means.tolist()),
|
||||
'var_stats': list(total_vars.tolist()),
|
||||
'total_frames': total_frames
|
||||
}
|
||||
with open(cmvn_file, 'w') as fout:
|
||||
fout.write(json.dumps(cmvn_info))
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
@@ -1,126 +0,0 @@
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
|
||||
import numpy as np
|
||||
import torchaudio
|
||||
import torchaudio.compliance.kaldi as kaldi
|
||||
import yaml
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="computer global cmvn",
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
|
||||
)
|
||||
parser.add_argument(
|
||||
"--dim",
|
||||
default=80,
|
||||
type=int,
|
||||
help="feature dimension",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--wav_path",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="the path of wav scps",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--config_file",
|
||||
type=str,
|
||||
help="the config file for computing cmvn",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--idx",
|
||||
default=1,
|
||||
required=True,
|
||||
type=int,
|
||||
help="index",
|
||||
)
|
||||
return parser
|
||||
|
||||
|
||||
def compute_fbank(wav_file,
|
||||
num_mel_bins=80,
|
||||
frame_length=25,
|
||||
frame_shift=10,
|
||||
dither=0.0,
|
||||
resample_rate=16000,
|
||||
speed=1.0,
|
||||
window_type="hamming"):
|
||||
waveform, sample_rate = torchaudio.load(wav_file)
|
||||
if resample_rate != sample_rate:
|
||||
waveform = torchaudio.transforms.Resample(orig_freq=sample_rate,
|
||||
new_freq=resample_rate)(waveform)
|
||||
if speed != 1.0:
|
||||
waveform, _ = torchaudio.sox_effects.apply_effects_tensor(
|
||||
waveform, resample_rate,
|
||||
[['speed', str(speed)], ['rate', str(resample_rate)]]
|
||||
)
|
||||
|
||||
waveform = waveform * (1 << 15)
|
||||
mat = kaldi.fbank(waveform,
|
||||
num_mel_bins=num_mel_bins,
|
||||
frame_length=frame_length,
|
||||
frame_shift=frame_shift,
|
||||
dither=dither,
|
||||
energy_floor=0.0,
|
||||
window_type=window_type,
|
||||
sample_frequency=resample_rate)
|
||||
|
||||
return mat.numpy()
|
||||
|
||||
|
||||
def main():
|
||||
parser = get_parser()
|
||||
args = parser.parse_args()
|
||||
|
||||
wav_scp_file = os.path.join(args.wav_path, "wav.{}.scp".format(args.idx))
|
||||
cmvn_file = os.path.join(args.wav_path, "cmvn.{}.json".format(args.idx))
|
||||
|
||||
mean_stats = np.zeros(args.dim)
|
||||
var_stats = np.zeros(args.dim)
|
||||
total_frames = 0
|
||||
|
||||
# with ReadHelper('ark:{}'.format(ark_file)) as ark_reader:
|
||||
# for key, mat in ark_reader:
|
||||
# mean_stats += np.sum(mat, axis=0)
|
||||
# var_stats += np.sum(np.square(mat), axis=0)
|
||||
# total_frames += mat.shape[0]
|
||||
|
||||
with open(args.config_file) as f:
|
||||
configs = yaml.safe_load(f)
|
||||
frontend_configs = configs.get("frontend_conf", {})
|
||||
num_mel_bins = frontend_configs.get("n_mels", 80)
|
||||
frame_length = frontend_configs.get("frame_length", 25)
|
||||
frame_shift = frontend_configs.get("frame_shift", 10)
|
||||
window_type = frontend_configs.get("window", "hamming")
|
||||
resample_rate = frontend_configs.get("fs", 16000)
|
||||
assert num_mel_bins == args.dim
|
||||
|
||||
with open(wav_scp_file) as f:
|
||||
lines = f.readlines()
|
||||
for line in lines:
|
||||
_, wav_file = line.strip().split()
|
||||
fbank = compute_fbank(wav_file,
|
||||
num_mel_bins=args.dim,
|
||||
frame_length=frame_length,
|
||||
frame_shift=frame_shift,
|
||||
resample_rate=resample_rate,
|
||||
window_type=window_type)
|
||||
mean_stats += np.sum(fbank, axis=0)
|
||||
var_stats += np.sum(np.square(fbank), axis=0)
|
||||
total_frames += fbank.shape[0]
|
||||
|
||||
cmvn_info = {
|
||||
'mean_stats': list(mean_stats.tolist()),
|
||||
'var_stats': list(var_stats.tolist()),
|
||||
'total_frames': total_frames
|
||||
}
|
||||
with open(cmvn_file, 'w') as fout:
|
||||
fout.write(json.dumps(cmvn_info))
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
@@ -1,39 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
# Begin configuration section.
|
||||
fbankdir=
|
||||
nj=32
|
||||
cmd=./utils/run.pl
|
||||
feats_dim=80
|
||||
config_file=
|
||||
scale=1.0
|
||||
|
||||
echo "$0 $@"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
# shellcheck disable=SC2046
|
||||
head -n $(awk -v lines="$(wc -l < ${fbankdir}/wav.scp)" -v scale="$scale" 'BEGIN { printf "%.0f\n", lines*scale }') ${fbankdir}/wav.scp > ${fbankdir}/wav.scp.scale
|
||||
|
||||
split_dir=${fbankdir}/cmvn/split_${nj};
|
||||
mkdir -p $split_dir
|
||||
split_scps=""
|
||||
for n in $(seq $nj); do
|
||||
split_scps="$split_scps $split_dir/wav.$n.scp"
|
||||
done
|
||||
utils/split_scp.pl ${fbankdir}/wav.scp.scale $split_scps || exit 1;
|
||||
|
||||
logdir=${fbankdir}/cmvn/log
|
||||
$cmd JOB=1:$nj $logdir/cmvn.JOB.log \
|
||||
python utils/compute_cmvn.py \
|
||||
--dim ${feats_dim} \
|
||||
--wav_path $split_dir \
|
||||
--config_file $config_file \
|
||||
--idx JOB \
|
||||
|
||||
python utils/combine_cmvn_file.py --dim ${feats_dim} --cmvn_dir $split_dir --nj $nj --output_dir ${fbankdir}/cmvn
|
||||
|
||||
python utils/cmvn_converter.py --cmvn_json ${fbankdir}/cmvn/cmvn.json --am_mvn ${fbankdir}/cmvn/am.mvn
|
||||
|
||||
echo "$0: Succeeded compute global cmvn"
|
||||
@@ -1,171 +0,0 @@
|
||||
from kaldiio import WriteHelper
|
||||
|
||||
import argparse
|
||||
import numpy as np
|
||||
import json
|
||||
import torch
|
||||
import torchaudio
|
||||
import torchaudio.compliance.kaldi as kaldi
|
||||
|
||||
|
||||
def compute_fbank(wav_file,
|
||||
num_mel_bins=80,
|
||||
frame_length=25,
|
||||
frame_shift=10,
|
||||
dither=0.0,
|
||||
resample_rate=16000,
|
||||
speed=1.0,
|
||||
window_type="hamming"):
|
||||
|
||||
waveform, sample_rate = torchaudio.load(wav_file)
|
||||
if resample_rate != sample_rate:
|
||||
waveform = torchaudio.transforms.Resample(orig_freq=sample_rate,
|
||||
new_freq=resample_rate)(waveform)
|
||||
if speed != 1.0:
|
||||
waveform, _ = torchaudio.sox_effects.apply_effects_tensor(
|
||||
waveform, resample_rate,
|
||||
[['speed', str(speed)], ['rate', str(resample_rate)]]
|
||||
)
|
||||
|
||||
waveform = waveform * (1 << 15)
|
||||
mat = kaldi.fbank(waveform,
|
||||
num_mel_bins=num_mel_bins,
|
||||
frame_length=frame_length,
|
||||
frame_shift=frame_shift,
|
||||
dither=dither,
|
||||
energy_floor=0.0,
|
||||
window_type=window_type,
|
||||
sample_frequency=resample_rate)
|
||||
|
||||
return mat.numpy()
|
||||
|
||||
|
||||
def get_parser():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="computer features",
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
|
||||
)
|
||||
parser.add_argument(
|
||||
"--wav-lists",
|
||||
"-w",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="input wav lists",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--text-files",
|
||||
"-t",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="input text files",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--dims",
|
||||
"-d",
|
||||
default=80,
|
||||
type=int,
|
||||
help="feature dims",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--max-lengths",
|
||||
"-m",
|
||||
default=1500,
|
||||
type=int,
|
||||
help="max frame numbers",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--sample-frequency",
|
||||
"-s",
|
||||
default=16000,
|
||||
type=int,
|
||||
help="sample frequency",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--speed-perturb",
|
||||
"-p",
|
||||
default="1.0",
|
||||
type=str,
|
||||
help="speed perturb",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--ark-index",
|
||||
"-a",
|
||||
default=1,
|
||||
required=True,
|
||||
type=int,
|
||||
help="ark index",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--output-dir",
|
||||
"-o",
|
||||
default=False,
|
||||
required=True,
|
||||
type=str,
|
||||
help="output dir",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--window-type",
|
||||
default="hamming",
|
||||
required=False,
|
||||
type=str,
|
||||
help="window type"
|
||||
)
|
||||
return parser
|
||||
|
||||
|
||||
def main():
|
||||
parser = get_parser()
|
||||
args = parser.parse_args()
|
||||
|
||||
ark_file = args.output_dir + "/ark/feats." + str(args.ark_index) + ".ark"
|
||||
scp_file = args.output_dir + "/ark/feats." + str(args.ark_index) + ".scp"
|
||||
text_file = args.output_dir + "/txt/text." + str(args.ark_index) + ".txt"
|
||||
feats_shape_file = args.output_dir + "/ark/len." + str(args.ark_index)
|
||||
text_shape_file = args.output_dir + "/txt/len." + str(args.ark_index)
|
||||
|
||||
ark_writer = WriteHelper('ark,scp:{},{}'.format(ark_file, scp_file))
|
||||
text_writer = open(text_file, 'w')
|
||||
feats_shape_writer = open(feats_shape_file, 'w')
|
||||
text_shape_writer = open(text_shape_file, 'w')
|
||||
|
||||
speed_perturb_list = args.speed_perturb.split(',')
|
||||
|
||||
for speed in speed_perturb_list:
|
||||
with open(args.wav_lists, 'r', encoding='utf-8') as wavfile:
|
||||
with open(args.text_files, 'r', encoding='utf-8') as textfile:
|
||||
for wav, text in zip(wavfile, textfile):
|
||||
s_w = wav.strip().split()
|
||||
wav_id = s_w[0]
|
||||
wav_file = s_w[1]
|
||||
|
||||
s_t = text.strip().split()
|
||||
text_id = s_t[0]
|
||||
txt = s_t[1:]
|
||||
fbank = compute_fbank(wav_file,
|
||||
num_mel_bins=args.dims,
|
||||
resample_rate=args.sample_frequency,
|
||||
speed=float(speed),
|
||||
window_type=args.window_type
|
||||
)
|
||||
feats_dims = fbank.shape[1]
|
||||
feats_lens = fbank.shape[0]
|
||||
if feats_lens >= args.max_lengths:
|
||||
continue
|
||||
txt_lens = len(txt)
|
||||
if speed == "1.0":
|
||||
wav_id_sp = wav_id
|
||||
else:
|
||||
wav_id_sp = wav_id + "_sp" + speed
|
||||
|
||||
feats_shape_writer.write(wav_id_sp + " " + str(feats_lens) + "," + str(feats_dims) + '\n')
|
||||
text_shape_writer.write(wav_id_sp + " " + str(txt_lens) + '\n')
|
||||
|
||||
text_writer.write(wav_id_sp + " " + " ".join(txt) + '\n')
|
||||
ark_writer(wav_id_sp, fbank)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
|
||||
@@ -1,54 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
. ./path.sh || exit 1;
|
||||
# Begin configuration section.
|
||||
nj=32
|
||||
cmd=./utils/run.pl
|
||||
|
||||
# feature configuration
|
||||
feats_dim=80
|
||||
sample_frequency=16000
|
||||
speed_perturb="1.0"
|
||||
window_type="hamming"
|
||||
max_lengths=1500
|
||||
|
||||
echo "$0 $@"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
data=$1
|
||||
logdir=$2
|
||||
fbankdir=$3
|
||||
|
||||
[ ! -f $data/wav.scp ] && echo "$0: no such file $data/wav.scp" && exit 1;
|
||||
[ ! -f $data/text ] && echo "$0: no such file $data/text" && exit 1;
|
||||
|
||||
python utils/split_data.py $data $data $nj
|
||||
|
||||
ark_dir=${fbankdir}/ark; mkdir -p ${ark_dir}
|
||||
text_dir=${fbankdir}/txt; mkdir -p ${text_dir}
|
||||
mkdir -p ${logdir}
|
||||
|
||||
$cmd JOB=1:$nj $logdir/make_fbank.JOB.log \
|
||||
python utils/compute_fbank.py -w $data/split${nj}/JOB/wav.scp -t $data/split${nj}/JOB/text \
|
||||
-d $feats_dim -s $sample_frequency -m ${max_lengths} -p ${speed_perturb} -a JOB -o ${fbankdir} \
|
||||
--window-type ${window_type} \
|
||||
|| exit 1;
|
||||
|
||||
for n in $(seq $nj); do
|
||||
cat ${ark_dir}/feats.$n.scp || exit 1
|
||||
done > $fbankdir/feats.scp || exit 1
|
||||
|
||||
for n in $(seq $nj); do
|
||||
cat ${text_dir}/text.$n.txt || exit 1
|
||||
done > $fbankdir/text || exit 1
|
||||
|
||||
for n in $(seq $nj); do
|
||||
cat ${ark_dir}/len.$n || exit 1
|
||||
done > $fbankdir/speech_shape || exit 1
|
||||
|
||||
for n in $(seq $nj); do
|
||||
cat ${text_dir}/len.$n || exit 1
|
||||
done > $fbankdir/text_shape || exit 1
|
||||
|
||||
echo "$0: Succeeded compute FBANK features"
|
||||
@@ -1,20 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
import argparse
|
||||
|
||||
from modelscope.pipelines import pipeline
|
||||
from modelscope.utils.constant import Tasks
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(
|
||||
description="download model configs",
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
|
||||
)
|
||||
parser.add_argument("--model_name",
|
||||
type=str,
|
||||
default="damo/speech_data2vec_pretrain-zh-cn-aishell2-16k-pytorch",
|
||||
help="model name in ModelScope")
|
||||
args = parser.parse_args()
|
||||
|
||||
inference_pipeline = pipeline(
|
||||
task=Tasks.auto_speech_recognition,
|
||||
model=args.model_name)
|
||||
@@ -1,370 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
# coding=utf8
|
||||
|
||||
# Copyright 2021 Jiayu DU
|
||||
|
||||
import sys
|
||||
import argparse
|
||||
import json
|
||||
import logging
|
||||
logging.basicConfig(stream=sys.stderr, level=logging.INFO, format='[%(levelname)s] %(message)s')
|
||||
|
||||
DEBUG = None
|
||||
|
||||
def GetEditType(ref_token, hyp_token):
|
||||
if ref_token == None and hyp_token != None:
|
||||
return 'I'
|
||||
elif ref_token != None and hyp_token == None:
|
||||
return 'D'
|
||||
elif ref_token == hyp_token:
|
||||
return 'C'
|
||||
elif ref_token != hyp_token:
|
||||
return 'S'
|
||||
else:
|
||||
raise RuntimeError
|
||||
|
||||
class AlignmentArc:
|
||||
def __init__(self, src, dst, ref, hyp):
|
||||
self.src = src
|
||||
self.dst = dst
|
||||
self.ref = ref
|
||||
self.hyp = hyp
|
||||
self.edit_type = GetEditType(ref, hyp)
|
||||
|
||||
def similarity_score_function(ref_token, hyp_token):
|
||||
return 0 if (ref_token == hyp_token) else -1.0
|
||||
|
||||
def insertion_score_function(token):
|
||||
return -1.0
|
||||
|
||||
def deletion_score_function(token):
|
||||
return -1.0
|
||||
|
||||
def EditDistance(
|
||||
ref,
|
||||
hyp,
|
||||
similarity_score_function = similarity_score_function,
|
||||
insertion_score_function = insertion_score_function,
|
||||
deletion_score_function = deletion_score_function):
|
||||
assert(len(ref) != 0)
|
||||
class DPState:
|
||||
def __init__(self):
|
||||
self.score = -float('inf')
|
||||
# backpointer
|
||||
self.prev_r = None
|
||||
self.prev_h = None
|
||||
|
||||
def print_search_grid(S, R, H, fstream):
|
||||
print(file=fstream)
|
||||
for r in range(R):
|
||||
for h in range(H):
|
||||
print(F'[{r},{h}]:{S[r][h].score:4.3f}:({S[r][h].prev_r},{S[r][h].prev_h}) ', end='', file=fstream)
|
||||
print(file=fstream)
|
||||
|
||||
R = len(ref) + 1
|
||||
H = len(hyp) + 1
|
||||
|
||||
# Construct DP search space, a (R x H) grid
|
||||
S = [ [] for r in range(R) ]
|
||||
for r in range(R):
|
||||
S[r] = [ DPState() for x in range(H) ]
|
||||
|
||||
# initialize DP search grid origin, S(r = 0, h = 0)
|
||||
S[0][0].score = 0.0
|
||||
S[0][0].prev_r = None
|
||||
S[0][0].prev_h = None
|
||||
|
||||
# initialize REF axis
|
||||
for r in range(1, R):
|
||||
S[r][0].score = S[r-1][0].score + deletion_score_function(ref[r-1])
|
||||
S[r][0].prev_r = r-1
|
||||
S[r][0].prev_h = 0
|
||||
|
||||
# initialize HYP axis
|
||||
for h in range(1, H):
|
||||
S[0][h].score = S[0][h-1].score + insertion_score_function(hyp[h-1])
|
||||
S[0][h].prev_r = 0
|
||||
S[0][h].prev_h = h-1
|
||||
|
||||
best_score = S[0][0].score
|
||||
best_state = (0, 0)
|
||||
|
||||
for r in range(1, R):
|
||||
for h in range(1, H):
|
||||
sub_or_cor_score = similarity_score_function(ref[r-1], hyp[h-1])
|
||||
new_score = S[r-1][h-1].score + sub_or_cor_score
|
||||
if new_score >= S[r][h].score:
|
||||
S[r][h].score = new_score
|
||||
S[r][h].prev_r = r-1
|
||||
S[r][h].prev_h = h-1
|
||||
|
||||
del_score = deletion_score_function(ref[r-1])
|
||||
new_score = S[r-1][h].score + del_score
|
||||
if new_score >= S[r][h].score:
|
||||
S[r][h].score = new_score
|
||||
S[r][h].prev_r = r - 1
|
||||
S[r][h].prev_h = h
|
||||
|
||||
ins_score = insertion_score_function(hyp[h-1])
|
||||
new_score = S[r][h-1].score + ins_score
|
||||
if new_score >= S[r][h].score:
|
||||
S[r][h].score = new_score
|
||||
S[r][h].prev_r = r
|
||||
S[r][h].prev_h = h-1
|
||||
|
||||
best_score = S[R-1][H-1].score
|
||||
best_state = (R-1, H-1)
|
||||
|
||||
if DEBUG:
|
||||
print_search_grid(S, R, H, sys.stderr)
|
||||
|
||||
# Backtracing best alignment path, i.e. a list of arcs
|
||||
# arc = (src, dst, ref, hyp, edit_type)
|
||||
# src/dst = (r, h), where r/h refers to search grid state-id along Ref/Hyp axis
|
||||
best_path = []
|
||||
r, h = best_state[0], best_state[1]
|
||||
prev_r, prev_h = S[r][h].prev_r, S[r][h].prev_h
|
||||
score = S[r][h].score
|
||||
# loop invariant:
|
||||
# 1. (prev_r, prev_h) -> (r, h) is a "forward arc" on best alignment path
|
||||
# 2. score is the value of point(r, h) on DP search grid
|
||||
while prev_r != None or prev_h != None:
|
||||
src = (prev_r, prev_h)
|
||||
dst = (r, h)
|
||||
if (r == prev_r + 1 and h == prev_h + 1): # Substitution or correct
|
||||
arc = AlignmentArc(src, dst, ref[prev_r], hyp[prev_h])
|
||||
elif (r == prev_r + 1 and h == prev_h): # Deletion
|
||||
arc = AlignmentArc(src, dst, ref[prev_r], None)
|
||||
elif (r == prev_r and h == prev_h + 1): # Insertion
|
||||
arc = AlignmentArc(src, dst, None, hyp[prev_h])
|
||||
else:
|
||||
raise RuntimeError
|
||||
best_path.append(arc)
|
||||
r, h = prev_r, prev_h
|
||||
prev_r, prev_h = S[r][h].prev_r, S[r][h].prev_h
|
||||
score = S[r][h].score
|
||||
|
||||
best_path.reverse()
|
||||
return (best_path, best_score)
|
||||
|
||||
def PrettyPrintAlignment(alignment, stream = sys.stderr):
|
||||
def get_token_str(token):
|
||||
if token == None:
|
||||
return "*"
|
||||
return token
|
||||
|
||||
def is_double_width_char(ch):
|
||||
if (ch >= '\u4e00') and (ch <= '\u9fa5'): # codepoint ranges for Chinese chars
|
||||
return True
|
||||
# TODO: support other double-width-char language such as Japanese, Korean
|
||||
else:
|
||||
return False
|
||||
|
||||
def display_width(token_str):
|
||||
m = 0
|
||||
for c in token_str:
|
||||
if is_double_width_char(c):
|
||||
m += 2
|
||||
else:
|
||||
m += 1
|
||||
return m
|
||||
|
||||
R = ' REF : '
|
||||
H = ' HYP : '
|
||||
E = ' EDIT : '
|
||||
for arc in alignment:
|
||||
r = get_token_str(arc.ref)
|
||||
h = get_token_str(arc.hyp)
|
||||
e = arc.edit_type if arc.edit_type != 'C' else ''
|
||||
|
||||
nr, nh, ne = display_width(r), display_width(h), display_width(e)
|
||||
n = max(nr, nh, ne) + 1
|
||||
|
||||
R += r + ' ' * (n-nr)
|
||||
H += h + ' ' * (n-nh)
|
||||
E += e + ' ' * (n-ne)
|
||||
|
||||
print(R, file=stream)
|
||||
print(H, file=stream)
|
||||
print(E, file=stream)
|
||||
|
||||
def CountEdits(alignment):
|
||||
c, s, i, d = 0, 0, 0, 0
|
||||
for arc in alignment:
|
||||
if arc.edit_type == 'C':
|
||||
c += 1
|
||||
elif arc.edit_type == 'S':
|
||||
s += 1
|
||||
elif arc.edit_type == 'I':
|
||||
i += 1
|
||||
elif arc.edit_type == 'D':
|
||||
d += 1
|
||||
else:
|
||||
raise RuntimeError
|
||||
return (c, s, i, d)
|
||||
|
||||
def ComputeTokenErrorRate(c, s, i, d):
|
||||
return 100.0 * (s + d + i) / (s + d + c)
|
||||
|
||||
def ComputeSentenceErrorRate(num_err_utts, num_utts):
|
||||
assert(num_utts != 0)
|
||||
return 100.0 * num_err_utts / num_utts
|
||||
|
||||
|
||||
class EvaluationResult:
|
||||
def __init__(self):
|
||||
self.num_ref_utts = 0
|
||||
self.num_hyp_utts = 0
|
||||
self.num_eval_utts = 0 # seen in both ref & hyp
|
||||
self.num_hyp_without_ref = 0
|
||||
|
||||
self.C = 0
|
||||
self.S = 0
|
||||
self.I = 0
|
||||
self.D = 0
|
||||
self.token_error_rate = 0.0
|
||||
|
||||
self.num_utts_with_error = 0
|
||||
self.sentence_error_rate = 0.0
|
||||
|
||||
def to_json(self):
|
||||
return json.dumps(self.__dict__)
|
||||
|
||||
def to_kaldi(self):
|
||||
info = (
|
||||
F'%WER {self.token_error_rate:.2f} [ {self.S + self.D + self.I} / {self.C + self.S + self.D}, {self.I} ins, {self.D} del, {self.S} sub ]\n'
|
||||
F'%SER {self.sentence_error_rate:.2f} [ {self.num_utts_with_error} / {self.num_eval_utts} ]\n'
|
||||
)
|
||||
return info
|
||||
|
||||
def to_sclite(self):
|
||||
return "TODO"
|
||||
|
||||
def to_espnet(self):
|
||||
return "TODO"
|
||||
|
||||
def to_summary(self):
|
||||
#return json.dumps(self.__dict__, indent=4)
|
||||
summary = (
|
||||
'==================== Overall Statistics ====================\n'
|
||||
F'num_ref_utts: {self.num_ref_utts}\n'
|
||||
F'num_hyp_utts: {self.num_hyp_utts}\n'
|
||||
F'num_hyp_without_ref: {self.num_hyp_without_ref}\n'
|
||||
F'num_eval_utts: {self.num_eval_utts}\n'
|
||||
F'sentence_error_rate: {self.sentence_error_rate:.2f}%\n'
|
||||
F'token_error_rate: {self.token_error_rate:.2f}%\n'
|
||||
F'token_stats:\n'
|
||||
F' - tokens:{self.C + self.S + self.D:>7}\n'
|
||||
F' - edits: {self.S + self.I + self.D:>7}\n'
|
||||
F' - cor: {self.C:>7}\n'
|
||||
F' - sub: {self.S:>7}\n'
|
||||
F' - ins: {self.I:>7}\n'
|
||||
F' - del: {self.D:>7}\n'
|
||||
'============================================================\n'
|
||||
)
|
||||
return summary
|
||||
|
||||
|
||||
class Utterance:
|
||||
def __init__(self, uid, text):
|
||||
self.uid = uid
|
||||
self.text = text
|
||||
|
||||
|
||||
def LoadUtterances(filepath, format):
|
||||
utts = {}
|
||||
if format == 'text': # utt_id word1 word2 ...
|
||||
with open(filepath, 'r', encoding='utf8') as f:
|
||||
for line in f:
|
||||
line = line.strip()
|
||||
if line:
|
||||
cols = line.split(maxsplit=1)
|
||||
assert(len(cols) == 2 or len(cols) == 1)
|
||||
uid = cols[0]
|
||||
text = cols[1] if len(cols) == 2 else ''
|
||||
if utts.get(uid) != None:
|
||||
raise RuntimeError(F'Found duplicated utterence id {uid}')
|
||||
utts[uid] = Utterance(uid, text)
|
||||
else:
|
||||
raise RuntimeError(F'Unsupported text format {format}')
|
||||
return utts
|
||||
|
||||
|
||||
def tokenize_text(text, tokenizer):
|
||||
if tokenizer == 'whitespace':
|
||||
return text.split()
|
||||
elif tokenizer == 'char':
|
||||
return [ ch for ch in ''.join(text.split()) ]
|
||||
else:
|
||||
raise RuntimeError(F'ERROR: Unsupported tokenizer {tokenizer}')
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser()
|
||||
# optional
|
||||
parser.add_argument('--tokenizer', choices=['whitespace', 'char'], default='whitespace', help='whitespace for WER, char for CER')
|
||||
parser.add_argument('--ref-format', choices=['text'], default='text', help='reference format, first col is utt_id, the rest is text')
|
||||
parser.add_argument('--hyp-format', choices=['text'], default='text', help='hypothesis format, first col is utt_id, the rest is text')
|
||||
# required
|
||||
parser.add_argument('--ref', type=str, required=True, help='input reference file')
|
||||
parser.add_argument('--hyp', type=str, required=True, help='input hypothesis file')
|
||||
|
||||
parser.add_argument('result_file', type=str)
|
||||
args = parser.parse_args()
|
||||
logging.info(args)
|
||||
|
||||
ref_utts = LoadUtterances(args.ref, args.ref_format)
|
||||
hyp_utts = LoadUtterances(args.hyp, args.hyp_format)
|
||||
|
||||
r = EvaluationResult()
|
||||
|
||||
# check valid utterances in hyp that have matched non-empty reference
|
||||
eval_utts = []
|
||||
r.num_hyp_without_ref = 0
|
||||
for uid in sorted(hyp_utts.keys()):
|
||||
if uid in ref_utts.keys(): # TODO: efficiency
|
||||
if ref_utts[uid].text.strip(): # non-empty reference
|
||||
eval_utts.append(uid)
|
||||
else:
|
||||
logging.warn(F'Found {uid} with empty reference, skipping...')
|
||||
else:
|
||||
logging.warn(F'Found {uid} without reference, skipping...')
|
||||
r.num_hyp_without_ref += 1
|
||||
|
||||
r.num_hyp_utts = len(hyp_utts)
|
||||
r.num_ref_utts = len(ref_utts)
|
||||
r.num_eval_utts = len(eval_utts)
|
||||
|
||||
with open(args.result_file, 'w+', encoding='utf8') as fo:
|
||||
for uid in eval_utts:
|
||||
ref = ref_utts[uid]
|
||||
hyp = hyp_utts[uid]
|
||||
|
||||
alignment, score = EditDistance(
|
||||
tokenize_text(ref.text, args.tokenizer),
|
||||
tokenize_text(hyp.text, args.tokenizer)
|
||||
)
|
||||
|
||||
c, s, i, d = CountEdits(alignment)
|
||||
utt_ter = ComputeTokenErrorRate(c, s, i, d)
|
||||
|
||||
# utt-level evaluation result
|
||||
print(F'{{"uid":{uid}, "score":{score}, "ter":{utt_ter:.2f}, "cor":{c}, "sub":{s}, "ins":{i}, "del":{d}}}', file=fo)
|
||||
PrettyPrintAlignment(alignment, fo)
|
||||
|
||||
r.C += c
|
||||
r.S += s
|
||||
r.I += i
|
||||
r.D += d
|
||||
|
||||
if utt_ter > 0:
|
||||
r.num_utts_with_error += 1
|
||||
|
||||
# corpus level evaluation result
|
||||
r.sentence_error_rate = ComputeSentenceErrorRate(r.num_utts_with_error, r.num_eval_utts)
|
||||
r.token_error_rate = ComputeTokenErrorRate(r.C, r.S, r.I, r.D)
|
||||
|
||||
print(r.to_summary(), file=fo)
|
||||
|
||||
print(r.to_json())
|
||||
print(r.to_kaldi())
|
||||
@@ -1,47 +0,0 @@
|
||||
from transformers import AutoTokenizer, AutoModel, pipeline
|
||||
import numpy as np
|
||||
import sys
|
||||
import os
|
||||
import torch
|
||||
from kaldiio import WriteHelper
|
||||
import re
|
||||
text_file_json = sys.argv[1]
|
||||
out_ark = sys.argv[2]
|
||||
out_scp = sys.argv[3]
|
||||
out_shape = sys.argv[4]
|
||||
device = int(sys.argv[5])
|
||||
model_path = sys.argv[6]
|
||||
|
||||
model = AutoModel.from_pretrained(model_path)
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
||||
extractor = pipeline(task="feature-extraction", model=model, tokenizer=tokenizer, device=device)
|
||||
|
||||
with open(text_file_json, 'r') as f:
|
||||
js = f.readlines()
|
||||
|
||||
|
||||
f_shape = open(out_shape, "w")
|
||||
with WriteHelper('ark,scp:{},{}'.format(out_ark, out_scp)) as writer:
|
||||
with torch.no_grad():
|
||||
for idx, line in enumerate(js):
|
||||
id, tokens = line.strip().split(" ", 1)
|
||||
tokens = re.sub(" ", "", tokens.strip())
|
||||
tokens = ' '.join([j for j in tokens])
|
||||
token_num = len(tokens.split(" "))
|
||||
outputs = extractor(tokens)
|
||||
outputs = np.array(outputs)
|
||||
embeds = outputs[0, 1:-1, :]
|
||||
|
||||
token_num_embeds, dim = embeds.shape
|
||||
if token_num == token_num_embeds:
|
||||
writer(id, embeds)
|
||||
shape_line = "{} {},{}\n".format(id, token_num_embeds, dim)
|
||||
f_shape.write(shape_line)
|
||||
else:
|
||||
print("{}, size has changed, {}, {}, {}".format(id, token_num, token_num_embeds, tokens))
|
||||
|
||||
|
||||
|
||||
f_shape.close()
|
||||
|
||||
|
||||
@@ -1,87 +0,0 @@
|
||||
#!/usr/bin/env perl
|
||||
# Copyright 2010-2012 Microsoft Corporation
|
||||
# Johns Hopkins University (author: Daniel Povey)
|
||||
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
# This script takes a list of utterance-ids or any file whose first field
|
||||
# of each line is an utterance-id, and filters an scp
|
||||
# file (or any file whose "n-th" field is an utterance id), printing
|
||||
# out only those lines whose "n-th" field is in id_list. The index of
|
||||
# the "n-th" field is 1, by default, but can be changed by using
|
||||
# the -f <n> switch
|
||||
|
||||
$exclude = 0;
|
||||
$field = 1;
|
||||
$shifted = 0;
|
||||
|
||||
do {
|
||||
$shifted=0;
|
||||
if ($ARGV[0] eq "--exclude") {
|
||||
$exclude = 1;
|
||||
shift @ARGV;
|
||||
$shifted=1;
|
||||
}
|
||||
if ($ARGV[0] eq "-f") {
|
||||
$field = $ARGV[1];
|
||||
shift @ARGV; shift @ARGV;
|
||||
$shifted=1
|
||||
}
|
||||
} while ($shifted);
|
||||
|
||||
if(@ARGV < 1 || @ARGV > 2) {
|
||||
die "Usage: filter_scp.pl [--exclude] [-f <field-to-filter-on>] id_list [in.scp] > out.scp \n" .
|
||||
"Prints only the input lines whose f'th field (default: first) is in 'id_list'.\n" .
|
||||
"Note: only the first field of each line in id_list matters. With --exclude, prints\n" .
|
||||
"only the lines that were *not* in id_list.\n" .
|
||||
"Caution: previously, the -f option was interpreted as a zero-based field index.\n" .
|
||||
"If your older scripts (written before Oct 2014) stopped working and you used the\n" .
|
||||
"-f option, add 1 to the argument.\n" .
|
||||
"See also: scripts/filter_scp.pl .\n";
|
||||
}
|
||||
|
||||
|
||||
$idlist = shift @ARGV;
|
||||
open(F, "<$idlist") || die "Could not open id-list file $idlist";
|
||||
while(<F>) {
|
||||
@A = split;
|
||||
@A>=1 || die "Invalid id-list file line $_";
|
||||
$seen{$A[0]} = 1;
|
||||
}
|
||||
|
||||
if ($field == 1) { # Treat this as special case, since it is common.
|
||||
while(<>) {
|
||||
$_ =~ m/\s*(\S+)\s*/ || die "Bad line $_, could not get first field.";
|
||||
# $1 is what we filter on.
|
||||
if ((!$exclude && $seen{$1}) || ($exclude && !defined $seen{$1})) {
|
||||
print $_;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
while(<>) {
|
||||
@A = split;
|
||||
@A > 0 || die "Invalid scp file line $_";
|
||||
@A >= $field || die "Invalid scp file line $_";
|
||||
if ((!$exclude && $seen{$A[$field-1]}) || ($exclude && !defined $seen{$A[$field-1]})) {
|
||||
print $_;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
# tests:
|
||||
# the following should print "foo 1"
|
||||
# ( echo foo 1; echo bar 2 ) | scripts/filter_scp.pl <(echo foo)
|
||||
# the following should print "bar 2".
|
||||
# ( echo foo 1; echo bar 2 ) | scripts/filter_scp.pl -f 2 <(echo 2)
|
||||
@@ -1,35 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
echo "$0 $@"
|
||||
data_dir=$1
|
||||
|
||||
if [ ! -f ${data_dir}/wav.scp ]; then
|
||||
echo "$0: wav.scp is not found"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ ! -f ${data_dir}/text ]; then
|
||||
echo "$0: text is not found"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
|
||||
|
||||
mkdir -p ${data_dir}/.backup
|
||||
|
||||
awk '{print $1}' ${data_dir}/wav.scp > ${data_dir}/.backup/wav_id
|
||||
awk '{print $1}' ${data_dir}/text > ${data_dir}/.backup/text_id
|
||||
|
||||
sort ${data_dir}/.backup/wav_id ${data_dir}/.backup/text_id | uniq -d > ${data_dir}/.backup/id
|
||||
|
||||
cp ${data_dir}/wav.scp ${data_dir}/.backup/wav.scp
|
||||
cp ${data_dir}/text ${data_dir}/.backup/text
|
||||
|
||||
mv ${data_dir}/wav.scp ${data_dir}/wav.scp.bak
|
||||
mv ${data_dir}/text ${data_dir}/text.bak
|
||||
|
||||
utils/filter_scp.pl -f 1 ${data_dir}/.backup/id ${data_dir}/wav.scp.bak | sort -k1,1 -u > ${data_dir}/wav.scp
|
||||
utils/filter_scp.pl -f 1 ${data_dir}/.backup/id ${data_dir}/text.bak | sort -k1,1 -u > ${data_dir}/text
|
||||
|
||||
rm ${data_dir}/wav.scp.bak
|
||||
rm ${data_dir}/text.bak
|
||||
@@ -1,52 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
echo "$0 $@"
|
||||
data_dir=$1
|
||||
|
||||
if [ ! -f ${data_dir}/feats.scp ]; then
|
||||
echo "$0: feats.scp is not found"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ ! -f ${data_dir}/text ]; then
|
||||
echo "$0: text is not found"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ ! -f ${data_dir}/speech_shape ]; then
|
||||
echo "$0: feature lengths is not found"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
if [ ! -f ${data_dir}/text_shape ]; then
|
||||
echo "$0: text lengths is not found"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
mkdir -p ${data_dir}/.backup
|
||||
|
||||
awk '{print $1}' ${data_dir}/feats.scp > ${data_dir}/.backup/wav_id
|
||||
awk '{print $1}' ${data_dir}/text > ${data_dir}/.backup/text_id
|
||||
|
||||
sort ${data_dir}/.backup/wav_id ${data_dir}/.backup/text_id | uniq -d > ${data_dir}/.backup/id
|
||||
|
||||
cp ${data_dir}/feats.scp ${data_dir}/.backup/feats.scp
|
||||
cp ${data_dir}/text ${data_dir}/.backup/text
|
||||
cp ${data_dir}/speech_shape ${data_dir}/.backup/speech_shape
|
||||
cp ${data_dir}/text_shape ${data_dir}/.backup/text_shape
|
||||
|
||||
mv ${data_dir}/feats.scp ${data_dir}/feats.scp.bak
|
||||
mv ${data_dir}/text ${data_dir}/text.bak
|
||||
mv ${data_dir}/speech_shape ${data_dir}/speech_shape.bak
|
||||
mv ${data_dir}/text_shape ${data_dir}/text_shape.bak
|
||||
|
||||
utils/filter_scp.pl -f 1 ${data_dir}/.backup/id ${data_dir}/feats.scp.bak | sort -k1,1 -u > ${data_dir}/feats.scp
|
||||
utils/filter_scp.pl -f 1 ${data_dir}/.backup/id ${data_dir}/text.bak | sort -k1,1 -u > ${data_dir}/text
|
||||
utils/filter_scp.pl -f 1 ${data_dir}/.backup/id ${data_dir}/speech_shape.bak | sort -k1,1 -u > ${data_dir}/speech_shape
|
||||
utils/filter_scp.pl -f 1 ${data_dir}/.backup/id ${data_dir}/text_shape.bak | sort -k1,1 -u > ${data_dir}/text_shape
|
||||
|
||||
rm ${data_dir}/feats.scp.bak
|
||||
rm ${data_dir}/text.bak
|
||||
rm ${data_dir}/speech_shape.bak
|
||||
rm ${data_dir}/text_shape.bak
|
||||
|
||||
@@ -1,22 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
|
||||
# Begin configuration section.
|
||||
nj=32
|
||||
cmd=./utils/run.pl
|
||||
|
||||
echo "$0 $@"
|
||||
|
||||
. utils/parse_options.sh || exit 1;
|
||||
|
||||
ark_dir=$1
|
||||
txt_dir=$2
|
||||
output_dir=$3
|
||||
|
||||
[ ! -d ${ark_dir}/ark ] && echo "$0: ark data is required" && exit 1;
|
||||
[ ! -d ${txt_dir}/txt ] && echo "$0: txt data is required" && exit 1;
|
||||
|
||||
for n in $(seq $nj); do
|
||||
echo "${ark_dir}/ark/feats.$n.ark ${txt_dir}/txt/text.$n.txt" || exit 1
|
||||
done > ${output_dir}/ark_txt.scp || exit 1
|
||||
|
||||
@@ -1,118 +0,0 @@
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
"--task",
|
||||
type=str,
|
||||
default="auto-speech-recognition",
|
||||
help="task name",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--type",
|
||||
type=str,
|
||||
default="generic-asr",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--am_model_name",
|
||||
type=str,
|
||||
default="model.pb",
|
||||
help="model file name",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--mode",
|
||||
type=str,
|
||||
default="paraformer",
|
||||
help="mode for decoding",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--lang",
|
||||
type=str,
|
||||
default="zh-cn",
|
||||
help="language",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--batch_size",
|
||||
type=int,
|
||||
default=1,
|
||||
help="batch size",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--am_model_config",
|
||||
type=str,
|
||||
default="config.yaml",
|
||||
help="config file",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--mvn_file",
|
||||
type=str,
|
||||
default="am.mvn",
|
||||
help="cmvn file",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--model_name",
|
||||
type=str,
|
||||
help="model name",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--pipeline_type",
|
||||
type=str,
|
||||
default="asr-inference",
|
||||
help="pipeline type",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--vocab_size",
|
||||
type=int,
|
||||
help="vocab_size",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--dataset",
|
||||
type=str,
|
||||
help="dataset name",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--output_dir",
|
||||
type=str,
|
||||
help="output path",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--nat",
|
||||
type=str,
|
||||
default="",
|
||||
help="nat",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--tag",
|
||||
type=str,
|
||||
default="exp1",
|
||||
help="model name tag",
|
||||
)
|
||||
args = parser.parse_args()
|
||||
|
||||
model = {
|
||||
"type": args.type,
|
||||
"am_model_name": args.am_model_name,
|
||||
"model_config": {
|
||||
"type": "pytorch",
|
||||
"code_base": "funasr",
|
||||
"mode": args.mode,
|
||||
"lang": args.lang,
|
||||
"batch_size": args.batch_size,
|
||||
"am_model_config": args.am_model_config,
|
||||
"mvn_file": args.mvn_file,
|
||||
"model": "speech_{}_asr{}-{}-16k-{}-vocab{}-pytorch-{}".format(args.model_name, args.nat, args.lang,
|
||||
args.dataset, args.vocab_size, args.tag),
|
||||
}
|
||||
}
|
||||
pipeline = {"type": args.pipeline_type}
|
||||
json_dict = {
|
||||
"framework": "pytorch",
|
||||
"task": args.task,
|
||||
"model": model,
|
||||
"pipeline": pipeline,
|
||||
}
|
||||
|
||||
with open(os.path.join(args.output_dir, "configuration.json"), "w") as f:
|
||||
json.dump(json_dict, f, indent=4)
|
||||
@@ -1,97 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
|
||||
# Copyright 2012 Johns Hopkins University (Author: Daniel Povey);
|
||||
# Arnab Ghoshal, Karel Vesely
|
||||
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
# MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
# See the Apache 2 License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
# Parse command-line options.
|
||||
# To be sourced by another script (as in ". parse_options.sh").
|
||||
# Option format is: --option-name arg
|
||||
# and shell variable "option_name" gets set to value "arg."
|
||||
# The exception is --help, which takes no arguments, but prints the
|
||||
# $help_message variable (if defined).
|
||||
|
||||
|
||||
###
|
||||
### The --config file options have lower priority to command line
|
||||
### options, so we need to import them first...
|
||||
###
|
||||
|
||||
# Now import all the configs specified by command-line, in left-to-right order
|
||||
for ((argpos=1; argpos<$#; argpos++)); do
|
||||
if [ "${!argpos}" == "--config" ]; then
|
||||
argpos_plus1=$((argpos+1))
|
||||
config=${!argpos_plus1}
|
||||
[ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
|
||||
. $config # source the config file.
|
||||
fi
|
||||
done
|
||||
|
||||
|
||||
###
|
||||
### Now we process the command line options
|
||||
###
|
||||
while true; do
|
||||
[ -z "${1:-}" ] && break; # break if there are no arguments
|
||||
case "$1" in
|
||||
# If the enclosing script is called with --help option, print the help
|
||||
# message and exit. Scripts should put help messages in $help_message
|
||||
--help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
|
||||
else printf "$help_message\n" 1>&2 ; fi;
|
||||
exit 0 ;;
|
||||
--*=*) echo "$0: options to scripts must be of the form --name value, got '$1'"
|
||||
exit 1 ;;
|
||||
# If the first command-line argument begins with "--" (e.g. --foo-bar),
|
||||
# then work out the variable name as $name, which will equal "foo_bar".
|
||||
--*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`;
|
||||
# Next we test whether the variable in question is undefned-- if so it's
|
||||
# an invalid option and we die. Note: $0 evaluates to the name of the
|
||||
# enclosing script.
|
||||
# The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
|
||||
# is undefined. We then have to wrap this test inside "eval" because
|
||||
# foo_bar is itself inside a variable ($name).
|
||||
eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
|
||||
|
||||
oldval="`eval echo \\$$name`";
|
||||
# Work out whether we seem to be expecting a Boolean argument.
|
||||
if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
|
||||
was_bool=true;
|
||||
else
|
||||
was_bool=false;
|
||||
fi
|
||||
|
||||
# Set the variable to the right value-- the escaped quotes make it work if
|
||||
# the option had spaces, like --cmd "queue.pl -sync y"
|
||||
eval $name=\"$2\";
|
||||
|
||||
# Check that Boolean-valued arguments are really Boolean.
|
||||
if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
|
||||
echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
|
||||
exit 1;
|
||||
fi
|
||||
shift 2;
|
||||
;;
|
||||
*) break;
|
||||
esac
|
||||
done
|
||||
|
||||
|
||||
# Check for an empty argument to the --cmd option, which can easily occur as a
|
||||
# result of scripting errors.
|
||||
[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
|
||||
|
||||
|
||||
true; # so this script returns exit code 0.
|
||||
@@ -1,19 +0,0 @@
|
||||
import os
|
||||
import shutil
|
||||
|
||||
from modelscope.hub.snapshot_download import snapshot_download
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
import sys
|
||||
|
||||
model = sys.argv[1]
|
||||
checkpoint_dir = sys.argv[2]
|
||||
checkpoint_name = sys.argv[3]
|
||||
|
||||
try:
|
||||
pretrained_model_path = snapshot_download(model, cache_dir=checkpoint_dir)
|
||||
except BaseException:
|
||||
raise BaseException(f"Please download pretrain model from ModelScope firstly.")
|
||||
shutil.copy(os.path.join(checkpoint_dir, checkpoint_name), os.path.join(pretrained_model_path, "model.pb"))
|
||||
shutil.copy(os.path.join(checkpoint_dir, "config.yaml"), os.path.join(pretrained_model_path, "config.yaml"))
|
||||
@@ -1,45 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
import sys
|
||||
|
||||
|
||||
def get_commandline_args(no_executable=True):
|
||||
extra_chars = [
|
||||
" ",
|
||||
";",
|
||||
"&",
|
||||
"|",
|
||||
"<",
|
||||
">",
|
||||
"?",
|
||||
"*",
|
||||
"~",
|
||||
"`",
|
||||
'"',
|
||||
"'",
|
||||
"\\",
|
||||
"{",
|
||||
"}",
|
||||
"(",
|
||||
")",
|
||||
]
|
||||
|
||||
# Escape the extra characters for shell
|
||||
argv = [
|
||||
arg.replace("'", "'\\''")
|
||||
if all(char not in arg for char in extra_chars)
|
||||
else "'" + arg.replace("'", "'\\''") + "'"
|
||||
for arg in sys.argv
|
||||
]
|
||||
|
||||
if no_executable:
|
||||
return " ".join(argv[1:])
|
||||
else:
|
||||
return sys.executable + " " + " ".join(argv)
|
||||
|
||||
|
||||
def main():
|
||||
print(get_commandline_args())
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -1,35 +0,0 @@
|
||||
from pathlib import Path
|
||||
|
||||
import torch
|
||||
import yaml
|
||||
|
||||
|
||||
class NoAliasSafeDumper(yaml.SafeDumper):
|
||||
# Disable anchor/alias in yaml because looks ugly
|
||||
def ignore_aliases(self, data):
|
||||
return True
|
||||
|
||||
|
||||
def yaml_no_alias_safe_dump(data, stream=None, **kwargs):
|
||||
"""Safe-dump in yaml with no anchor/alias"""
|
||||
return yaml.dump(
|
||||
data, stream, allow_unicode=True, Dumper=NoAliasSafeDumper, **kwargs
|
||||
)
|
||||
|
||||
|
||||
def gen_conf(file, out_dir):
|
||||
conf = torch.load(file)["config"]
|
||||
conf["oss_bucket"] = "null"
|
||||
print(conf)
|
||||
output_dir = Path(out_dir)
|
||||
output_dir.mkdir(parents=True, exist_ok=True)
|
||||
with (output_dir / "config.yaml").open("w", encoding="utf-8") as f:
|
||||
yaml_no_alias_safe_dump(conf, f, indent=4, sort_keys=False)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import sys
|
||||
|
||||
in_f = sys.argv[1]
|
||||
out_f = sys.argv[2]
|
||||
gen_conf(in_f, out_f)
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user