vad

游雁

2024-02-27 1b21c1120c447f2d6f012b00bb94ea97bfc9f92c

 README.md

@@ -3,11 +3,10 @@
([简体中文](./README_zh.md)|English)

# FunASR: A Fundamental End-to-End Speech Recognition Toolkit
<p align="left">
    <a href=""><img src="https://img.shields.io/badge/OS-Linux%2C%20Win%2C%20Mac-brightgreen.svg"></a>
    <a href=""><img src="https://img.shields.io/badge/Python->=3.7,<=3.10-aff.svg"></a>
    <a href=""><img src="https://img.shields.io/badge/Pytorch-%3E%3D1.11-blue"></a>
</p>


[![PyPI](https://img.shields.io/pypi/v/funasr)](https://pypi.org/project/funasr/)


<strong>FunASR</strong> hopes to build a bridge between academic research and industrial applications on speech recognition. By supporting the training & finetuning of the industrial-grade speech recognition model, researchers and developers can conduct research and production of speech recognition models more conveniently, and promote the development of speech recognition ecology. ASR for Fun！

@@ -27,7 +26,15 @@


<a name="whats-new"></a>
## What's new: 
## What's new:
- 2024/01/30：funasr-1.0 has been released ([docs](https://github.com/alibaba-damo-academy/FunASR/discussions/1319))
- 2024/01/30：emotion recognition models are new supported. [model link](https://www.modelscope.cn/models/iic/emotion2vec_base_finetuned/summary), modified from [repo](https://github.com/ddlBoJack/emotion2vec).
- 2024/01/25: Offline File Transcription Service 4.2, Offline File Transcription Service of English 1.3 released，optimized the VAD (Voice Activity Detection) data processing method, significantly reducing peak memory usage, memory leak optimization; Real-time Transcription Service 1.7 released，optimizatized the client-side；([docs](runtime/readme.md))
- 2024/01/09: The Funasr SDK for Windows version 2.0 has been released, featuring support for The offline file transcription service (CPU) of Mandarin 4.1, The offline file transcription service (CPU) of English 1.2, The real-time transcription service (CPU) of Mandarin 1.6. For more details, please refer to the official documentation or release notes([FunASR-Runtime-Windows](https://www.modelscope.cn/models/damo/funasr-runtime-win-cpu-x64/summary))
- 2024/01/03: File Transcription Service 4.0 released, Added support for 8k models, optimized timestamp mismatch issues and added sentence-level timestamps, improved the effectiveness of English word FST hotwords, supported automated configuration of thread parameters, and fixed known crash issues as well as memory leak problems, refer to ([docs](runtime/readme.md#file-transcription-service-mandarin-cpu)).
- 2024/01/03: Real-time Transcription Service 1.6 released，The 2pass-offline mode supports Ngram language model decoding and WFST hotwords, while also addressing known crash issues and memory leak problems, ([docs](runtime/readme.md#the-real-time-transcription-service-mandarin-cpu))
- 2024/01/03: Fixed known crash issues as well as memory leak problems, ([docs](runtime/readme.md#file-transcription-service-english-cpu)).
- 2023/12/04: The Funasr SDK for Windows version 1.0 has been released, featuring support for The offline file transcription service (CPU) of Mandarin, The offline file transcription service (CPU) of English, The real-time transcription service (CPU) of Mandarin. For more details, please refer to the official documentation or release notes([FunASR-Runtime-Windows](https://www.modelscope.cn/models/damo/funasr-runtime-win-cpu-x64/summary))
- 2023/11/08: The offline file transcription service 3.0 (CPU) of Mandarin has been released, adding punctuation large model, Ngram language model, and wfst hot words. For detailed information, please refer to [docs](runtime#file-transcription-service-mandarin-cpu). 
- 2023/10/17: The offline file transcription service (CPU) of English has been released. For more details, please refer to ([docs](runtime#file-transcription-service-english-cpu)).
- 2023/10/13: [SlideSpeech](https://slidespeech.github.io/): A large scale multi-modal audio-visual corpus with a significant amount of real-time synchronized slides.
@@ -42,7 +49,19 @@
<a name="Installation"></a>
## Installation

Please ref to [installation docs](https://alibaba-damo-academy.github.io/FunASR/en/installation/installation.html)
```shell
pip3 install -U funasr
```
Or install from source code
``` sh
git clone https://github.com/alibaba/FunASR.git && cd FunASR
pip3 install -e ./
```
Install modelscope for the pretrained models (Optional)

```shell
pip3 install -U modelscope
```

## Model Zoo
FunASR has open-sourced a large number of pre-trained models on industrial data. You are free to use, copy, modify, and share FunASR models under the [Model License Agreement](./MODEL_LICENSE). Below are some representative models, for more models please refer to the [Model Zoo]().
@@ -50,17 +69,16 @@
(Note: 🤗 represents the Huggingface model zoo link, ⭐ represents the ModelScope model zoo link)


|                                                                           Model Name                                                                            |                                Task Details                                 |          Training Date           | Parameters |
|:---------------------------------------------------------------------------------------------------------------------------------------------------------------:|:---------------------------------------------------------------------------:|:--------------------------------:|:----------:|
|     paraformer-zh ([⭐](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary)  [🤗]() )     |             speech recognition, with timestamps, non-streaming              |      60000 hours, Mandarin       |    220M    |
|                 paraformer-zh-spk ( [⭐](https://modelscope.cn/models/damo/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn/summary)  [🤗]() )                 | speech recognition with speaker diarization, with timestamps, non-streaming |      60000 hours, Mandarin       |    220M    |
|        paraformer-zh-online ( [⭐](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online/summary) [🤗]() )         |                      speech recognition, non-streaming                      |      60000 hours, Mandarin       |    220M    |
|          paraformer-en ( [⭐](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020/summary) [🤗]() )          |             speech recognition, with timestamps, non-streaming              |       50000 hours, English       |    220M    |
|                                                                paraformer-en-spk ([🤗]() [⭐]() )                                                                |         speech recognition with speaker diarization, non-streaming          |       50000 hours, English       |    220M    |
|                      conformer-en ( [⭐](https://modelscope.cn/models/damo/speech_conformer_asr-en-16k-vocab4199-pytorch/summary) [🤗]() )                       |                      speech recognition, non-streaming                      |       50000 hours, English       |    220M    |
|                      ct-punc ( [⭐](https://modelscope.cn/models/damo/punc_ct-transformer_cn-en-common-vocab471067-large/summary) [🤗]() )                       |                           punctuation restoration                           |    100M, Mandarin and English    |    1.1G    | 
|                           fsmn-vad ( [⭐](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary) [🤗]() )                           |                          voice activity detection                           | 5000 hours, Mandarin and English |    0.4M    | 
|                           fa-zh ( [⭐](https://modelscope.cn/models/damo/speech_timestamp_prediction-v1-16k-offline/summary) [🤗]() )                            |                            timestamp prediction                             |       5000 hours, Mandarin       |    38M     | 
|                                                                                                         Model Name                                                                                                         |                    Task Details                    |          Training Data           | Parameters |
|:--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|:--------------------------------------------------:|:--------------------------------:|:----------:|
|          paraformer-zh <br> ([⭐](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary)  [🤗](https://huggingface.co/funasr/paraformer-tp) )           | speech recognition, with timestamps, non-streaming |      60000 hours, Mandarin       |    220M    |
| <nobr>paraformer-zh-streaming <br> ( [⭐](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online/summary) [🤗](https://huggingface.co/funasr/paraformer-zh-streaming) )</nobr> |           speech recognition, streaming            |      60000 hours, Mandarin       |    220M    |
|               paraformer-en <br> ( [⭐](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020/summary) [🤗](https://huggingface.co/funasr/paraformer-en) )                | speech recognition, with timestamps, non-streaming |       50000 hours, English       |    220M    |
|                            conformer-en <br> ( [⭐](https://modelscope.cn/models/damo/speech_conformer_asr-en-16k-vocab4199-pytorch/summary) [🤗](https://huggingface.co/funasr/conformer-en) )                             |         speech recognition, non-streaming          |       50000 hours, English       |    220M    |
|                               ct-punc <br> ( [⭐](https://modelscope.cn/models/damo/punc_ct-transformer_cn-en-common-vocab471067-large/summary) [🤗](https://huggingface.co/funasr/ct-punc) )                               |              punctuation restoration               |    100M, Mandarin and English    |    1.1G    | 
|                                   fsmn-vad <br> ( [⭐](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary) [🤗](https://huggingface.co/funasr/fsmn-vad) )                                   |              voice activity detection              | 5000 hours, Mandarin and English |    0.4M    | 
|                                     fa-zh <br> ( [⭐](https://modelscope.cn/models/damo/speech_timestamp_prediction-v1-16k-offline/summary) [🤗](https://huggingface.co/funasr/fa-zh) )                                     |                timestamp prediction                |       5000 hours, Mandarin       |    38M     | 
|                                       cam++ <br> ( [⭐](https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common/summary) [🤗](https://huggingface.co/funasr/campplus) )                                        |        speaker verification/diarization            |            5000 hours            |    7.2M    | 



@@ -71,47 +89,121 @@

<a name="quick-start"></a>
## Quick Start
Quick start for new users（[tutorial](https://alibaba-damo-academy.github.io/FunASR/en/funasr/quick_start.html)）

FunASR supports inference and fine-tuning of models trained on industrial data for tens of thousands of hours. For more details, please refer to [modelscope_egs](https://alibaba-damo-academy.github.io/FunASR/en/modelscope_pipeline/quick_start.html). It also supports training and fine-tuning of models on academic standard datasets. For more information, please refer to [egs](https://alibaba-damo-academy.github.io/FunASR/en/academic_recipe/asr_recipe.html).

Below is a quick start tutorial. Test audio files ([Mandarin](https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/vad_example.wav), [English]()).

### Command-line usage

```shell
funasr ++model=paraformer-zh ++vad_model="fsmn-vad" ++punc_model="ct-punc" ++input=asr_example_zh.wav
```

Notes: Support recognition of single audio file, as well as file list in Kaldi-style wav.scp format: `wav_id wav_pat`

### Speech Recognition (Non-streaming)
```python
from funasr import infer

p = infer(model="paraformer-zh", vad_model="fsmn-vad", punc_model="ct-punc", model_hub="ms")

res = p("asr_example_zh.wav", batch_size_token=5000)
from funasr import AutoModel
# paraformer-zh is a multi-functional asr model
# use vad, punc, spk or not as you need
model = AutoModel(model="paraformer-zh", model_revision="v2.0.4",
                  vad_model="fsmn-vad", vad_model_revision="v2.0.4",
                  punc_model="ct-punc-c", punc_model_revision="v2.0.4",
                  # spk_model="cam++", spk_model_revision="v2.0.2",
                  )
res = model.generate(input=f"{model.model_path}/example/asr_example.wav", 
                     batch_size_s=300, 
                     hotword='魔搭')
print(res)
```
Note: `model_hub`: represents the model repository, `ms` stands for selecting ModelScope download, `hf` stands for selecting Huggingface download.

### Speech Recognition (Streaming)
```python
from funasr import infer

p = infer(model="paraformer-zh-streaming", model_hub="ms")
from funasr import AutoModel

chunk_size = [0, 10, 5] #[0, 10, 5] 600ms, [0, 8, 4] 480ms
param_dict = {"cache": dict(), "is_final": False, "chunk_size": chunk_size, "encoder_chunk_look_back": 4, "decoder_chunk_look_back": 1}
encoder_chunk_look_back = 4 #number of chunks to lookback for encoder self-attention
decoder_chunk_look_back = 1 #number of encoder chunks to lookback for decoder cross-attention

import torchaudio
speech = torchaudio.load("asr_example_zh.wav")[0][0]
speech_length = speech.shape[0]
model = AutoModel(model="paraformer-zh-streaming", model_revision="v2.0.4")

stride_size = chunk_size[1] * 960
sample_offset = 0
for sample_offset in range(0, speech_length, min(stride_size, speech_length - sample_offset)):
    param_dict["is_final"] = True if sample_offset + stride_size >= speech_length - 1 else False
    input = speech[sample_offset: sample_offset + stride_size]
    rec_result = p(input=input, param_dict=param_dict)
    print(rec_result)
import soundfile
import os

wav_file = os.path.join(model.model_path, "example/asr_example.wav")
speech, sample_rate = soundfile.read(wav_file)
chunk_stride = chunk_size[1] * 960 # 600ms

cache = {}
total_chunk_num = int(len((speech)-1)/chunk_stride+1)
for i in range(total_chunk_num):
    speech_chunk = speech[i*chunk_stride:(i+1)*chunk_stride]
    is_final = i == total_chunk_num - 1
    res = model.generate(input=speech_chunk, cache=cache, is_final=is_final, chunk_size=chunk_size, encoder_chunk_look_back=encoder_chunk_look_back, decoder_chunk_look_back=decoder_chunk_look_back)
    print(res)
```
Note: `chunk_size` is the configuration for streaming latency.` [0,10,5]` indicates that the real-time display granularity is `10*60=600ms`, and the lookahead information is `5*60=300ms`. Each inference input is `600ms` (sample points are `16000*0.6=960`), and the output is the corresponding text. For the last speech segment input, `is_final=True` needs to be set to force the output of the last word.

Quick start for new users can be found in [docs](https://alibaba-damo-academy.github.io/FunASR/en/funasr/quick_start_zh.html)
### Voice Activity Detection (Non-Streaming)
```python
from funasr import AutoModel

model = AutoModel(model="fsmn-vad", model_revision="v2.0.4")
wav_file = f"{model.model_path}/example/asr_example.wav"
res = model.generate(input=wav_file)
print(res)
```
Note: The output format of the VAD model is: `[[beg1, end1], [beg2, end2], ..., [begN, endN]]`, where `begN/endN` indicates the starting/ending point of the `N-th` valid audio segment, measured in milliseconds.

### Voice Activity Detection (Streaming)
```python
from funasr import AutoModel

chunk_size = 200 # ms
model = AutoModel(model="fsmn-vad", model_revision="v2.0.4")

import soundfile

wav_file = f"{model.model_path}/example/vad_example.wav"
speech, sample_rate = soundfile.read(wav_file)
chunk_stride = int(chunk_size * sample_rate / 1000)

cache = {}
total_chunk_num = int(len((speech)-1)/chunk_stride+1)
for i in range(total_chunk_num):
    speech_chunk = speech[i*chunk_stride:(i+1)*chunk_stride]
    is_final = i == total_chunk_num - 1
    res = model.generate(input=speech_chunk, cache=cache, is_final=is_final, chunk_size=chunk_size)
    if len(res[0]["value"]):
        print(res)
```
Note: The output format for the streaming VAD model can be one of four scenarios:
- `[[beg1, end1], [beg2, end2], .., [begN, endN]]`：The same as the offline VAD output result mentioned above.
- `[[beg, -1]]`：Indicates that only a starting point has been detected.
- `[[-1, end]]`：Indicates that only an ending point has been detected.
- `[]`：Indicates that neither a starting point nor an ending point has been detected. 

The output is measured in milliseconds and represents the absolute time from the starting point.
### Punctuation Restoration
```python
from funasr import AutoModel

model = AutoModel(model="ct-punc", model_revision="v2.0.4")
res = model.generate(input="那今天的会就到这里吧 happy new year 明年见")
print(res)
```
### Timestamp Prediction
```python
from funasr import AutoModel

model = AutoModel(model="fa-zh", model_revision="v2.0.4")
wav_file = f"{model.model_path}/example/asr_example.wav"
text_file = f"{model.model_path}/example/text.txt"
res = model.generate(input=(wav_file, text_file), data_type=("sound", "text"))
print(res)
```

More examples ref to [docs](https://github.com/alibaba-damo-academy/FunASR/tree/main/examples/industrial_data_pretraining)

[//]: # (FunASR supports inference and fine-tuning of models trained on industrial datasets of tens of thousands of hours. For more details, please refer to &#40;[modelscope_egs]&#40;https://alibaba-damo-academy.github.io/FunASR/en/modelscope_pipeline/quick_start.html&#41;&#41;. It also supports training and fine-tuning of models on academic standard datasets. For more details, please refer to&#40;[egs]&#40;https://alibaba-damo-academy.github.io/FunASR/en/academic_recipe/asr_recipe.html&#41;&#41;. The models include speech recognition &#40;ASR&#41;, speech activity detection &#40;VAD&#41;, punctuation recovery, language model, speaker verification, speaker separation, and multi-party conversation speech recognition. For a detailed list of models, please refer to the [Model Zoo]&#40;https://github.com/alibaba-damo-academy/FunASR/blob/main/docs/model_zoo/modelscope_models.md&#41;:)

@@ -126,7 +218,7 @@
For more detailed information, please refer to the [service deployment documentation](runtime/readme.md).


<a name="Community Communication"></a>
<a name="contact"></a>
## Community Communication
If you encounter problems in use, you can directly raise Issues on the github page.

@@ -138,13 +230,13 @@

## Contributors

| <div align="left"><img src="docs/images/nwpu.png" width="260"/> | <img src="docs/images/China_Telecom.png" width="200"/> </div>  | <img src="docs/images/RapidAI.png" width="200"/> </div> | <img src="docs/images/aihealthx.png" width="200"/> </div> | <img src="docs/images/XVERSE.png" width="250"/> </div> |
|:---------------------------------------------------------------:|:--------------------------------------------------------------:|:-------------------------------------------------------:|:-----------------------------------------------------------:|:------------------------------------------------------:|
| <div align="left"><img src="docs/images/alibaba.png" width="260"/> | <div align="left"><img src="docs/images/nwpu.png" width="260"/> | <img src="docs/images/China_Telecom.png" width="200"/> </div>  | <img src="docs/images/RapidAI.png" width="200"/> </div> | <img src="docs/images/aihealthx.png" width="200"/> </div> | <img src="docs/images/XVERSE.png" width="250"/> </div> |
|:------------------------------------------------------------------:|:---------------------------------------------------------------:|:--------------------------------------------------------------:|:-------------------------------------------------------:|:-----------------------------------------------------------:|:------------------------------------------------------:|

The contributors can be found in [contributors list](./Acknowledge.md)

## License
This project is licensed under the [The MIT License](https://opensource.org/licenses/MIT). FunASR also contains various third-party components and some code modified from other repos under other open source licenses.
This project is licensed under [The MIT License](https://opensource.org/licenses/MIT). FunASR also contains various third-party components and some code modified from other repos under other open source licenses.
The use of pretraining model is subject to [model license](./MODEL_LICENSE)


@@ -164,10 +256,16 @@
}
@inproceedings{gao22b_interspeech,
  author={Zhifu Gao and ShiLiang Zhang and Ian McLoughlin and Zhijie Yan},
  title={{Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition}},
  title={Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition},
  year=2022,
  booktitle={Proc. Interspeech 2022},
  pages={2063--2067},
  doi={10.21437/Interspeech.2022-9996}
}
@inproceedings{shi2023seaco,
  author={Xian Shi and Yexin Yang and Zerui Li and Yanni Chen and Zhifu Gao and Shiliang Zhang},
  title={SeACo-Paraformer: A Non-Autoregressive ASR System with Flexible and Effective Hotword Customization Ability},
  year={2023},
  booktitle={ICASSP2024}
}
```