README.md

---
license: mit
language:
- en
- zh
base_model:
- CosyVoice2
pipeline_tag: text-to-speech
library_name: transformers
tags:
- CosyVoice2
- Speech
---

# CosyVoice2  
This version of CosyVoice2 has been converted to run on the Axera NPU using **w8a16** quantization.
Compatible with Pulsar2 version: 4.2

## Convert tools links:
For those who are interested in model conversion, you can try to export axmodel through the original repo : 
[Cosyvoice](https://github.com/FunAudioLLM/CosyVoice)

[Pulsar2 Link, How to Convert LLM from Huggingface to axmodel](https://pulsar2-docs.readthedocs.io/en/latest/appendix/build_llm.html) 

[AXera NPU HOST LLM Runtime](https://github.com/AXERA-TECH/Cosyvoice2.Axera) 

## Support Platform

- AX650
  - AX650N DEMO Board
  - [M4N-Dock(爱芯派Pro)](https://wiki.sipeed.com/hardware/zh/maixIV/m4ndock/m4ndock.html)
  - [M.2 Accelerator card](https://axcl-docs.readthedocs.io/zh-cn/latest/doc_guide_hardware.html)

**Speech Generation**  
| Stage | Time |
|------|------|
| llm prefill ( input_token_num + prompt_token_num 在 [0,128 ] ) | 104 ms  | 
| llm prefill ( input_token_num + prompt_token_num 在 [128,256 ] ) | 234 ms  | 
| Decode  |  21.24 token/s |

## How to use

Download all files from this repository to the device  

### 1. PrePare  

#### 1.1 Copy this project to AX650 Board  

#### 1.2 Prepare Dependencies  

**Running HTTP Tokenizer Server** and **Processing Prompt Speech** require these Python packages. If you run these two step on a PC, install them on the PC.  
```
pip3 install -r scripts/requirements.txt
```  

### 2. Start HTTP Tokenizer Server  
```
cd scripts
python cosyvoice2_tokenizer.py --host {your host} --port {your port}   
```


### 3. Run on Axera Device
There are 2 kinds of device, AX650 Board , AXCL aarch64 Board and AXCL x86 Board.

#### 3.1 Run on AX650 Board  
1) Moidfy the HTTP host in `run_ax650.sh`.  

2) Run `run_ax650.sh`  
```shell
root@ax650 ~/Cosyvoice2 # bash run_ax650.sh 
rm: cannot remove 'output*.wav': No such file or directory
[I][                            Init][ 108]: LLM init start
[I][                            Init][  34]: connect http://10.122.86.184:12345 ok
bos_id: 0, eos_id: 1773
  7% | ███                               |   2 /  27 [3.11s<42.04s, 0.64 count/s] embed_selector init ok[I][                            Init][ 138]: attr.axmodel_num:24
100% | ████████████████████████████████ |  27 /  27 [10.32s<10.32s, 2.62 count/s] init post axmodel ok,remain_cmm(7178 MB)
[I][                            Init][ 216]: max_token_len : 1023
[I][                            Init][ 221]: kv_cache_size : 128, kv_cache_num: 1023
[I][                            Init][ 229]: prefill_token_num : 128
[I][                            Init][ 233]: grp: 1, prefill_max_token_num : 1
[I][                            Init][ 233]: grp: 2, prefill_max_token_num : 128
[I][                            Init][ 233]: grp: 3, prefill_max_token_num : 256
[I][                            Init][ 233]: grp: 4, prefill_max_token_num : 384
[I][                            Init][ 233]: grp: 5, prefill_max_token_num : 512
[I][                            Init][ 237]: prefill_max_token_num : 512
[I][                            Init][ 249]: LLM init ok
[I][                            Init][ 154]: Token2Wav init ok
[I][                            main][ 273]: 
[I][                             Run][ 388]: input token num : 142, prefill_split_num : 2
[I][                             Run][ 422]: input_num_token:128
[I][                             Run][ 422]: input_num_token:14
[I][                             Run][ 607]: ttft: 236.90 ms
[Main/Token2Wav Thread] Processing batch of 28 tokens...
Successfully saved audio to output_0.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 53 tokens...
Successfully saved audio to output_1.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 78 tokens...
Successfully saved audio to output_2.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 78 tokens...
Successfully saved audio to output_3.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 78 tokens...
Successfully saved audio to output_4.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 78 tokens...
Successfully saved audio to output_5.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 78 tokens...
Successfully saved audio to output_6.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 78 tokens...
Successfully saved audio to output_7.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 78 tokens...
Successfully saved audio to output_8.wav (32-bit Float PCM).
[Main/Token2Wav Thread] Processing batch of 78 tokens...
Successfully saved audio to output_9.wav (32-bit Float PCM).
[I][                             Run][ 723]: hit eos, llm finished
[I][                             Run][ 753]: llm finished
[Main/Token2Wav Thread] Buffer is empty and LLM finished. Exiting.


[I][                             Run][ 758]: total decode tokens:271
[N][                             Run][ 759]: hit eos,avg 21.47 token/s

Successfully saved audio to output_10.wav (32-bit Float PCM).
Successfully saved audio to output.wav (32-bit Float PCM).

Voice generation pipeline completed.
Type "q" to exit, Ctrl+c to stop current running
text >> 
```

Output Speech：
[output.wav](asset/output.wav)


####  Or run on AX650 Board with Gradio GUI 
1) Start server
```
bash run_api_ax650.sh
```
2) Start Gradio GUI
```
python scripts/gradio_demo.py
```

#### 3.2 Run on AXCL aarch64 Board
```
bash run_axcl_aarch64.sh
```
#### Or run on AXCL aarch64 Board with Gradio GUI 
1) Start server
```
bash run_api_axcl_aarch64.sh
```
2) Start Gradio GUI
```
python scripts/gradio_demo.py
```
3) Open the page from a browser  
The page url is : `https://{your device ip}:7860`  

Note that you need to run these two commands in the project root directory.

#### 3.3 Run on AXCL x86 Board
```
bash run_axcl_x86.sh
```
#### Or run on AXCL aarch64 Board with Gradio GUI 
1) Start server
```
bash run_api_axcl_x86.sh
```
2) Start Gradio GUI
```
python scripts/gradio_demo.py
```
3) Open the page from a browser  
The page url is : `https://{your device ip}:7860`  

Note that you need to run these two commands in the project root directory.

![](./gradio.png)

### Optional. Process Prompt Speech    
If you want to replicate a specific sound, do this step.  
You can use audio in asset/ .  

#### (1). Downlaod wetext  
```
pip3 install modelscope
modelscope download --model pengzhendong/wetext --local_dir pengzhendong/wetext
```

#### (2). Process Prompt Speech  
Example:
```
python3 scripts/process_prompt.py --prompt_text  asset/zh_man1.txt --prompt_speech asset/zh_man1.wav --output zh_man1
```

Pass parameters according to the actual situation.
```
python3 scripts/process_prompt.py -h

usage: process_prompt.py [-h] [--model_dir MODEL_DIR] [--wetext_dir WETEXT_DIR] [--sample_rate SAMPLE_RATE] [--prompt_text PROMPT_TEXT] [--prompt_speech PROMPT_SPEECH]
                         [--output OUTPUT]

options:
  -h, --help            show this help message and exit
  --model_dir MODEL_DIR
                        tokenizer configuration directionary
  --wetext_dir WETEXT_DIR
                        path to wetext
  --sample_rate SAMPLE_RATE
                        Sampling rate for prompt audio
  --prompt_text PROMPT_TEXT
                        The text content of the prompt(reference) audio. Text or file path.
  --prompt_speech PROMPT_SPEECH
                        The path to prompt(reference) audio.
  --output OUTPUT       Output data storage directory
```

After executing the above command, files like the following will be generated:
```
flow_embedding.txt  
flow_prompt_speech_token.txt  
llm_embedding.txt  
llm_prompt_speech_token.txt  
prompt_speech_feat.txt  
prompt_text.txt
```

When you run run_ax650.sh, pass the output path here to the prompt_files parameter of the run_ax650.sh script.