mirror of
https://github.com/modelscope/FunASR.git
synced 2026-10-12 02:19:18 +00:00
zhifu gaoandLauraGPT
4dc8e5b4f0
Signed-off-by: LauraGPT <[email protected]> Co-authored-by: LauraGPT <[email protected]>
59 lines
2.5 KiB
Markdown
59 lines
2.5 KiB
Markdown
# ONNX Runtime binary JSONL output
|
|
|
|
The native ONNX Runtime binaries log results by default. Add `--output-format jsonl` when another process needs one machine-readable record for every completed input.
|
|
|
|
## Offline recognition
|
|
|
|
From `runtime/onnxruntime`, run:
|
|
|
|
```shell
|
|
./build/bin/funasr-onnx-offline \
|
|
--model-dir /path/to/timestamp-capable-model \
|
|
--wav-path /path/to/audio.wav \
|
|
--output-format jsonl \
|
|
> results.jsonl 2> runtime.log
|
|
```
|
|
|
|
The same option is available on `funasr-onnx-offline-rtf`.
|
|
|
|
## Two-pass recognition
|
|
|
|
```shell
|
|
./build/bin/funasr-onnx-2pass \
|
|
--model-dir /path/to/offline-model \
|
|
--online-model-dir /path/to/online-model \
|
|
--vad-dir /path/to/vad-model \
|
|
--punc-dir /path/to/online-punctuation-model \
|
|
--wav-path /path/to/audio.wav \
|
|
--mode 2pass \
|
|
--output-format jsonl \
|
|
> results.jsonl 2> runtime.log
|
|
```
|
|
|
|
The same option is available on `funasr-onnx-2pass-rtf`. Valid modes are `offline`, `online`, and `2pass`. The native two-pass runtime requires both a VAD model and an online punctuation model for successful inference.
|
|
|
|
## Record schema
|
|
|
|
Each stdout line is one JSON object:
|
|
|
|
```json
|
|
{"key":"utt-001","mode":"offline","stamp_sents":[{"end":920,"start":0,"text_seg":"hello world"}],"text":"hello world","timestamp":[[0,420],[460,920]]}
|
|
```
|
|
|
|
| Field | Type | Meaning |
|
|
| --- | --- | --- |
|
|
| `key` | string | The first column of a `wav.scp` row, or `wav_default_id` for a direct file input. |
|
|
| `mode` | string | `offline`, `online`, or `2pass`. |
|
|
| `text` | string | The completed transcript. In `2pass` mode this is the corrected offline transcript. |
|
|
| `timestamp` | array | The native token/word timestamp array returned by the model. |
|
|
| `stamp_sents` | array | Native sentence timestamp objects when the binary and model provide them. Two-pass binaries currently return an empty array here. |
|
|
|
|
Use a timestamp-capable model to receive timestamp values. If a model returns no timestamp, or the native payload is empty, malformed, or not a JSON array, the corresponding field is `[]`.
|
|
|
|
## Streaming and batch behavior
|
|
|
|
- JSON records are written only to stdout. Runtime diagnostics and existing progress logs remain on stderr.
|
|
- A `wav.scp` input produces one line per successfully completed key.
|
|
- RTF binaries protect each complete line from interleaving, but worker completion order is not guaranteed. Join records by `key` instead of relying on line order.
|
|
- Omitting `--output-format` is equivalent to `--output-format log` and preserves the existing logging behavior.
|