---
license: mit
language:
- en
pipeline_tag: automatic-speech-recognition
base_model:
- FunAudioLLM/SenseVoiceSmall
---
# SenseVoice
FunASR SenseVoice on Axera, official repo: https://github.com/FunAudioLLM/SenseVoice

## TODO

- [x] 支持 AX630C
- [x] 支持 C++
- [x] 支持 FastAPI

## 功能
 - 语音识别
 - 自动识别语言(支持中文、英文、粤语、日语、韩语)
 - 情感识别
 - 自动标点
 - 支持流式识别
 
## 支持平台

- [x] AX650N
- [x] AX630C

## Table of contents

- [环境安装](#环境安装)
- [使用](#使用)
- [准确率](#准确率)
- [技术讨论](#技术讨论)


## 环境安装

Python==3.12

```
sudo apt-get install libsndfile-dev
pip install -r requirements.txt
```


####  安装 pyaxenigne

参考 https://github.com/AXERA-TECH/pyaxengine 安装 NPU Python API

在 0.1.3rc2 上测试通过，可通过
```
pip install https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3.rc2/axengine-0.1.3-py3-none-any.whl
```

安装，或把版本号更改为你想使用的版本

## 使用

### Python

```
cd python
python3 main.py --input ../example/en.mp3
[INFO] Available providers:  ['AxEngineExecutionProvider']
{'input': '../example/en.mp3', 'language': 'auto', 'streaming': False}
......
RTF: 0.036785734138361184    Latency: 0.2639744281768799s  Total length: 7.176s
ASR result: the tribal chieftain called for the boy and presented him with fifty pieces of gold
```

运行参数说明:  
| 参数名称 | 说明 | 默认值 |
| --- | --- | --- |
| --input/-i | 输入音频文件 | |
| --language/-l | 识别语言，支持auto, zh, en, yue, ja, ko | auto |
| --streaming | 流式识别 | |

### CPP

 - AX650
```
./cpp/ax650/test_sensevoice -a example/zh.mp3 -p sensevoice_ax650/
Init asr success, take 0.2130seconds
Result: 开饭时间早上九点至下午五点
RTF(0.21 / 5.62) = 0.0372
```

 - AX630C
```
./cpp/ax630c/test_sensevoice -a example/zh.mp3 -p sensevoice_ax630c/
```

对应的源码在[Github](https://github.com/AXERA-TECH/ax_asr_api)上


### 示例  
example下有测试音频  

如 中文测试
```
cd python
python main.py -i example/zh.mp3
```
输出
```
RTF: 0.04386647134764582    Latency: 0.2463541030883789s  Total length: 5.616s
ASR result: 开饭时间早上九点至下午五点

```

流式识别  

```
python main.py -i example/zh.mp3 --streaming
```
输出
```
{'timestamps': [540], 'text': '开'}
{'timestamps': [540, 780, 1080], 'text': '开放时'}
{'timestamps': [540, 780, 1080, 1260, 1740], 'text': '开放时间早'}
{'timestamps': [540, 780, 1080, 1260, 1740, 1920, 2340], 'text': '开放时间早上9'}
{'timestamps': [540, 780, 1080, 1260, 1740, 1920, 2340, 2640], 'text': '开放时间早上9点'}
{'timestamps': [540, 780, 1080, 1260, 1740, 1920, 2340, 2640, 3060], 'text': '开放时间早上9点至'}
{'timestamps': [540, 780, 1080, 1260, 1740, 1920, 2340, 2640, 3060, 3780, 4020], 'text': '开放时间早上9点至下午'}
{'timestamps': [540, 780, 1080, 1260, 1740, 1920, 2340, 2640, 3060, 3780, 4020, 4440, 4620], 'text': '开放时间早上9点至下午五点'}
RTF: 0.03678379235444246

```

### Gradio DEMO

```
cd python
python3 gradio_demo.py
[INFO] Available providers:  ['AxEngineExecutionProvider']
[INFO] Using provider: AxEngineExecutionProvider
[INFO] Chip type: ChipType.MC50
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Engine version: 2.12.0s
[INFO] Model type: 2 (triple core)
[INFO] Compiler version: 5.0 76f70fdc
* Running on local URL:  https://xxx.xxx.xxx.xxx:7861
* Running on local URL:  https://172.18.0.1:7861
* Running on local URL:  https://172.17.0.1:7861
* Running on local URL:  https://0.0.0.0:7861
* To create a public link, set `share=True` in `launch()`.
```


![DEMO_Gradio](https://cdn-uploads.huggingface.co/production/uploads/660d19e9955bacf5a22a9b7b/ClusyqVUEm_gXTfvST7dg.png)


## 准确率

使用WER(Word-Error-Rate)作为评价标准  

**WER = 2.0%**  

### 复现测试结果

```
./download_datasets.sh
python test_wer.py -d aishell -g datasets/ground_truth.txt --language zh
```

## 技术讨论

- Github issues
- QQ 群: 139953715