--- 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.