Document the current sm_86 Windows CUDA prebuilt limit and direct RTX 50 / Blackwell users to CPU binaries or source builds with CMAKE_CUDA_ARCHITECTURES=120 until a dedicated asset is available.
Link the v1.3.19 README news item directly to long-session diagnostics and correct the repository roles page to show #3101 is still awaiting reporter retest logs.
Add benchmark RTF / RTFx reproducibility notes and link them from README, vLLM guides, and docs index.
Documents the public vLLM benchmark scope, the separate website H100 benchmark scope, and required fields for comparable reports.
The old comparison column cherry-picked the best cell from different models
(170x / emotion / 17x-CPU are SenseVoice; vLLM is Fun-ASR-Nano; streaming is
Paraformer), implying one model does all of it. Reframe as a toolkit comparison:
lead with the flagship Fun-ASR-Nano, label each capability with the model that
provides it, and use the flagship 340x realtime for top speed.
The header said "170x faster than Whisper", but 170x is realtime speed; per the
benchmark table FunASR is ~13-26x faster than Whisper. Use the flagship
Fun-ASR-Nano numbers consistent with the benchmark: up to 340x realtime, 26x
faster than Whisper-large-v3. Applied across en/zh/ja/ko.
The Usage streaming snippet used input="chunk.wav" (a file that does not
exist) and a single one-shot generate() call missing is_final /
encoder_chunk_look_back / the chunk loop, so it neither ran nor
demonstrated streaming.
Replace it with the real chunk-by-chunk loop (matching the repo example
examples/industrial_data_pretraining/paraformer_streaming/demo.py):
read audio, iterate fixed-stride chunks, pass cache + is_final +
look-back, print partial text per chunk.
Verified on GPU (paraformer-zh-streaming): emits incremental text per
chunk and reconstructs the full sentence.
Co-authored-by: LauraGPT <[email protected]>
The quickstart used input="meeting.wav", a file that does not exist, so
the very first example failed for every new user. It also showed a
hand-written [time] Speaker N: text block that model.generate() never
produces (the API returns raw SenseVoice tags in result[0]["text"] and
per-utterance speaker/timestamps in result[0]["sentence_info"]).
- Use the canonical FunASR sample URL (already used across the examples)
so the first run downloads and transcribes with no setup.
- Show the small loop that turns sentence_info into the labeled output,
and the real output for that sample.
- Replace the remaining meeting.wav placeholders with the same URL.
Verified end-to-end on GPU (SenseVoice + VAD + cam++).
Co-authored-by: LauraGPT <[email protected]>
- New funasr CLI with argparse: positional audio, --model, --language, --output-format (text/json/srt/tsv), --spk, --timestamps
- Old Hydra CLI preserved as funasr-hydra
- Updated README.md and README_zh.md with CLI examples
FsmnVADStreaming.forward() is inference-only (no loss computation),
so standard train_ds.py training does not work. Removing the
finetune.sh and docs that were incorrectly added.
The parameter tuning guide (vad_kwargs) remains valid and is
documented in the issue thread (#2961).
LauraGPT