返回模型

说明文档

VoxCPM

此版本的 VoxCPM 已转换为在 Axera NPU 上运行，采用 w8a16 量化。兼容 Pulsar2 版本：4.2

转换工具链接：

对于对模型转换感兴趣的用户，可以通过原始仓库尝试导出 axmodel： VoxCPM 官方

Pulsar2 链接，如何将 Huggingface 的 LLM 转换为 axmodel

AXera NPU HOST LLM 运行时

支持平台

AX650
- AX650N DEMO 开发板
- M4N-Dock（爱芯派Pro）
- M.2 加速卡

使用方法

将此仓库中的所有文件下载到设备上

1. 运行 Python 示例

1. 安装 voxcpm axinfer 包

git clone -b 1.0.4-axmode_infer https://github.com/techshoww/VoxCPM.git 
cd VoxCPM
pip3 install .

2. 下载 zipenhancer

pip3 install modelscope  
modelscope download --model iic/speech_zipenhancer_ans_multiloss_16k_base --local_dir iic/speech_zipenhancer_ans_multiloss_16k_base

3. 在 Axera 设备上运行

进入本项目的根目录，运行：

python3 run_ax650.py

2. 运行 C++ 示例

1. 安装 transformers

pip3 install transformers>=4.56.2

2. 启动 tokenizer 服务

python3 tokenizer.py --port 9999

3. 运行 C++ 示例

bash run_ax650.sh

AXERA-TECH/VoxCPM

作者 AXERA-TECH

text-to-speech transformers

↓ 1 ♥ 0

创建时间: 2025-09-30 07:05:00+00:00

更新时间: 2025-10-20 10:52:58+00:00

在 Hugging Face 上查看

文件 (98)

.gitattributes

README.md

VoxCPM-0.5B/config.json

VoxCPM-0.5B/special_tokens_map.json

VoxCPM-0.5B/tokenizer.json

VoxCPM-0.5B/tokenizer_config.json

assets/en_man1.mp3

assets/en_man1.txt

assets/en_woman1.mp3

assets/en_woman1.txt

assets/zh_man1.txt

assets/zh_man1.wav

assets/zh_man2.mp3

assets/zh_man2.txt

assets/zh_woman1.txt

assets/zh_woman1.wav

axmodels/audio_vae.decoder.onnx ONNX

axmodels/audio_vae.encoder.onnx ONNX

axmodels/enc_to_lm_proj.axmodel

axmodels/feat_encoder.in_proj.axmodel

axmodels/feat_encoder.in_proj.onnx ONNX

axmodels/feat_encoder.special_token.npy

axmodels/feat_encoder.special_token.txt

axmodels/fsq_layer.axmodel

axmodels/lm_to_dit_proj.axmodel

axmodels/locdit.part1.axmodel

axmodels/locdit.part3.axmodel

axmodels/rand_noise.txt

axmodels/res_to_dit_proj.axmodel

axmodels/stop_predictor.axmodel

axmodels/stop_predictor.onnx ONNX

base_lm-axmodels/MiniCPMForCausalLM_p64_l0_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l10_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l11_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l12_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l13_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l14_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l15_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l16_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l17_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l18_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l19_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l1_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l20_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l21_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l22_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l23_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l2_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l3_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l4_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l5_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l6_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l7_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l8_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_p64_l9_together.axmodel

base_lm-axmodels/MiniCPMForCausalLM_post.axmodel

base_lm-axmodels/model.embed_tokens.weight.bfloat16.bin

base_lm-axmodels/model.embed_tokens.weight.float32.bin

base_lm-axmodels/model.embed_tokens.weight.npy

config.json

feat_decoder_estimator_decoder-axmodels/MiniCPMForCausalLM_p5_l0_together.axmodel

feat_decoder_estimator_decoder-axmodels/MiniCPMForCausalLM_p5_l1_together.axmodel

feat_decoder_estimator_decoder-axmodels/MiniCPMForCausalLM_p5_l2_together.axmodel

feat_decoder_estimator_decoder-axmodels/MiniCPMForCausalLM_p5_l3_together.axmodel

feat_decoder_estimator_decoder-axmodels/MiniCPMForCausalLM_post.axmodel

feat_decoder_estimator_decoder-axmodels/model.embed_tokens.weight.bfloat16.bin

feat_decoder_estimator_decoder-axmodels/model.embed_tokens.weight.float32.bin

feat_decoder_estimator_decoder-axmodels/model.embed_tokens.weight.npy

feat_encoder_encoder-axmodels/MiniCPMForCausalLM_p3_l0_together.axmodel

feat_encoder_encoder-axmodels/MiniCPMForCausalLM_p3_l1_together.axmodel

feat_encoder_encoder-axmodels/MiniCPMForCausalLM_p3_l2_together.axmodel

feat_encoder_encoder-axmodels/MiniCPMForCausalLM_p3_l3_together.axmodel

feat_encoder_encoder-axmodels/MiniCPMForCausalLM_post.axmodel

feat_encoder_encoder-axmodels/model.embed_tokens.weight.bfloat16.bin

feat_encoder_encoder-axmodels/model.embed_tokens.weight.float32.bin

feat_encoder_encoder-axmodels/model.embed_tokens.weight.npy

main_ax650

onnxruntime-aarch64-none-gnu-1.16.0/lib/cmake/onnxruntime/onnxruntimeConfig.cmake

onnxruntime-aarch64-none-gnu-1.16.0/lib/cmake/onnxruntime/onnxruntimeConfigVersion.cmake

onnxruntime-aarch64-none-gnu-1.16.0/lib/cmake/onnxruntime/onnxruntimeTargets-release.cmake

onnxruntime-aarch64-none-gnu-1.16.0/lib/cmake/onnxruntime/onnxruntimeTargets.cmake

onnxruntime-aarch64-none-gnu-1.16.0/lib/libonnxruntime.so

onnxruntime-aarch64-none-gnu-1.16.0/lib/libonnxruntime.so.1.16.0

onnxruntime-aarch64-none-gnu-1.16.0/lib/libonnxruntime_providers_shared.so

onnxruntime-aarch64-none-gnu-1.16.0/lib/pkgconfig/libonnxruntime.pc

residual_lm-axmodels/MiniCPMForCausalLM_p64_l0_together.axmodel

residual_lm-axmodels/MiniCPMForCausalLM_p64_l1_together.axmodel

residual_lm-axmodels/MiniCPMForCausalLM_p64_l2_together.axmodel

residual_lm-axmodels/MiniCPMForCausalLM_p64_l3_together.axmodel

residual_lm-axmodels/MiniCPMForCausalLM_p64_l4_together.axmodel

residual_lm-axmodels/MiniCPMForCausalLM_p64_l5_together.axmodel

residual_lm-axmodels/MiniCPMForCausalLM_post.axmodel

residual_lm-axmodels/model.embed_tokens.weight.bfloat16.bin

residual_lm-axmodels/model.embed_tokens.weight.float32.bin

residual_lm-axmodels/model.embed_tokens.weight.npy

run_ax650.py

run_ax650.sh

tokenizer.py