AXERA-TECH / chatterbox-onestep

huggingface.co
Total runs: 43
24-hour runs: 1
7-day runs: -3
30-day runs: -198
Model's Last Updated: August 21 2026
text-to-speech

Introduction of chatterbox-onestep

Model Details of chatterbox-onestep

chatterbox-s3gen-onestep AXMODEL

精度 cosine ≈ 0.9932 | 推理耗时 181.08 ms | 模型大小 124.7 MB | 实时率 RTF ≈ 0.018

源码 / 复现

想要复现导出与编译流程、查看或修改 Python/C++ SDK 源码(含 OpenAI-Compatible server/client)、 运行 DSP 自测?源码仓库: https://github.com/ml-inory/chatterbox-onestep

  • model_convert/ :ONNX 导出 + Pulsar2 编译脚本(可复现本 AXMODEL)
  • python/ / cpp/ :完整 SDK 源码与示例
  • cpp/tests/ :FFT/Griffin-Lim 本地自测
性能与实时率(RTF)
场景 耗时 RTF
AX650C 板端 NPU(本模型,单步) 181 ms / 512 帧(10.24s 音频) ≈ 0.018
NVIDIA L4 GPU(同模型,单步) 47.2 ms ≈ 0.0046
教师 10 步(L4,参考) ~540 ms/句 ≈ 0.06

RTF = 生成耗时 / 音频时长,< 1 即快于实时。本模型 S3Gen 单步在板端生成 10 秒音频的 mel 只需约 0.18 秒;相比教师 10 步,单步化带来约 10 倍加速(RTF 0.06 → 0.018,同板卡口径)。

快速开始(只需两步)
1. 安装环境
bash setup.sh
2. 跑推理
bash run.sh
目录说明
目录 用途
models/ AXMODEL 模型文件 + 元信息
python/ Python SDK(pyaxengine)
cpp/ C++ SDK(AX Engine runtime)
model_convert/ 模型导出 & 编译脚本(可复现)
reports/ 各阶段报告
setup.sh 一键安装依赖
run.sh 一键运行推理
常见问题

Q: import 报错找不到 pyaxengine? A: 运行 bash setup.sh 会自动安装。

Q: 怎么在自己的代码里用? A: 参考 python/demo.py ,核心就 3 行:

from chatterbox-s3gen-onestep_sdk import ModelSDK
sdk = ModelSDK("models/model.axmodel")
result = sdk.run(your_input)

Q: 想自己重新编译? A: 进入 model_convert/ ,确保 Pulsar2 可用后运行 bash compile_pulsar2.sh 。 原始编译使用 Docker 镜像 docker-registry.aitsw.axera-tech.com/pulsar2:7.0 。

示例音频(学生单步 vs 教师 10 步,同噪声 A/B)

三句话的端到端合成示例: 学生 1 步(本模型) 与 教师 10 步(官方) 对比,均为 24kHz WAV、响度归一化。

句 1:Technology is best when it brings people together.

句 2:The quick brown fox jumps over the lazy dog.

句 3:A journey of a thousand miles begins with a single step.

OpenAI-Compatible 服务(板端 torch-free)

提供 OpenAI 风格的 POST /v1/audio/speech (另含 GET /v1/models ):输入 S3 speech token 序列, 输出 wav( response_format=wav )或 mel( response_format=mel )。 Python 与 C++ 两版均可在 AX 板上运行, 不依赖 torch / ffmpeg / FFTW ;文本 → token 的 T3 环节在宿主(torch)完成后调用本服务。

Python 版(依赖 numpy + pyaxengine)
python3 python/openai_server.py --model models/model.axmodel --port 8000
python3 python/openai_client.py --tokens-npy sample_input/tokens.npy --out out.wav
C++ 版(自实现 Bluestein FFT + Griffin-Lim,仅依赖 AX Engine)

本仓库 cpp/bin/ 提供 aarch64 预编译产物(Release + fast-math);C++ 完整源码见 GitHub: https://github.com/ml-inory/chatterbox-onestep

mkdir -p cpp/build && cd cpp/build
cmake .. -DCMAKE_TOOLCHAIN_FILE=${AX_RUNTIME_ROOT}/toolchain.cmake -DAX_RUNTIME_ROOT=${AX_RUNTIME_ROOT}
make -j$(nproc)
./openai_server --model ../../models/model.axmodel --assets ../assets --port 8000
./openai_client --url http://127.0.0.1:8000/v1/audio/speech --tokens-file <tokens_int32.bin> --out out.wav

不想自己编译也可以直接用 cpp/bin/ 里的预编译产物(板端加 LD_LIBRARY_PATH=/soc/lib ):

./cpp/bin/openai_server --model models/model.axmodel --assets cpp/assets --port 8000

说明:

  • 板端实测:Python wav 请求约 10s(numpy Griffin-Lim),C++ wav 请求约 23s(20 次迭代);
  • cpp/assets/ 提供 mel_inv_24k.bin (线性谱逆矩阵)与 default_embedding.bin (内置音色);
  • DSP 核心见 cpp/include/dsp.hpp ,本地自测见 cpp/tests/gl_test.cpp (FFT 对 numpy 误差约 1e-5)。
音色克隆(Voice Cloning,embedding 级)

本模型以 192 维 xvector 音色 embedding 为条件,支持"参考音频 → 换音色"的 embedding 级克隆 (完整官方克隆还含 10s 参考 prompt 条件,当前板端 AXMODEL 未导出该路径,需要可另行定制)。

步骤 1(宿主,torch) :从参考音频提取 embedding

python3 python/extract_voice_embedding.py --wav ref.wav \
    --ckpt-dir /path/to/chatterbox_models --out ref_embedding.npy

步骤 2(板端) :调用 OpenAI 服务时带上 voice.embedding

curl -X POST http://<board>:8000/v1/audio/speech -H 'Content-Type: application/json' \
  -d "$(python3 - <<'PY'
import json, numpy as np
emb = np.load('ref_embedding.npy').reshape(-1).tolist()
print(json.dumps({"input": [12,34,56], "voice": {"embedding": emb}, "response_format": "wav"}))
PY
)"

input 为 S3 speech tokens(文本 → token 的 T3 在宿主完成);也可在 python/openai_client.py 基础上扩展,直接读 ref_embedding.npy 填 voice.embedding 。

Runs of AXERA-TECH chatterbox-onestep on huggingface.co

43
Total runs
1
24-hour runs
3
3-day runs
-3
7-day runs
-198
30-day runs

More Information About chatterbox-onestep huggingface.co Model

More chatterbox-onestep license Visit here:

https://choosealicense.com/licenses/apache-2.0

chatterbox-onestep huggingface.co

chatterbox-onestep huggingface.co is an AI model on huggingface.co that provides chatterbox-onestep's model effect (), which can be used instantly with this AXERA-TECH chatterbox-onestep model. huggingface.co supports a free trial of the chatterbox-onestep model, and also provides paid use of the chatterbox-onestep. Support call chatterbox-onestep model through api, including Node.js, Python, http.

chatterbox-onestep huggingface.co Url

https://huggingface.co/AXERA-TECH/chatterbox-onestep

AXERA-TECH chatterbox-onestep online free

chatterbox-onestep huggingface.co is an online trial and call api platform, which integrates chatterbox-onestep's modeling effects, including api services, and provides a free online trial of chatterbox-onestep, you can try chatterbox-onestep online for free by clicking the link below.

AXERA-TECH chatterbox-onestep online free url in huggingface.co:

https://huggingface.co/AXERA-TECH/chatterbox-onestep

chatterbox-onestep install

chatterbox-onestep is an open source model from GitHub that offers a free installation service, and any user can find chatterbox-onestep on GitHub to install. At the same time, huggingface.co provides the effect of chatterbox-onestep install, users can directly use chatterbox-onestep installed effect in huggingface.co for debugging and trial. It also supports api for free installation.

chatterbox-onestep install url in huggingface.co:

https://huggingface.co/AXERA-TECH/chatterbox-onestep

Url of chatterbox-onestep

chatterbox-onestep huggingface.co Url

Provider of chatterbox-onestep huggingface.co

AXERA-TECH
ORGANIZATIONS

Other API from AXERA-TECH

huggingface.co

Total runs: 930
Run Growth: 921
Growth Rate: 100.00%
Updated:September 24 2026
huggingface.co

Total runs: 120
Run Growth: 50
Growth Rate: 41.67%
Updated:October 08 2026
huggingface.co

Total runs: 62
Run Growth: 22
Growth Rate: 34.92%
Updated:October 12 2026
huggingface.co

Total runs: 56
Run Growth: 26
Growth Rate: 49.06%
Updated:October 12 2026
huggingface.co

Total runs: 49
Run Growth: 14
Growth Rate: 28.57%
Updated:October 12 2026
huggingface.co

Total runs: 40
Run Growth: 53
Growth Rate: 84.13%
Updated:January 08 2026
huggingface.co

Total runs: 36
Run Growth: 2
Growth Rate: 6.06%
Updated:August 14 2026
huggingface.co

Total runs: 35
Run Growth: 11
Growth Rate: 28.95%
Updated:May 07 2026
huggingface.co

Total runs: 35
Run Growth: 5
Growth Rate: 15.15%
Updated:October 12 2026
huggingface.co

Total runs: 28
Run Growth: 19
Growth Rate: 67.86%
Updated:May 15 2026
huggingface.co

Total runs: 28
Run Growth: 7
Growth Rate: 24.14%
Updated:July 31 2026
huggingface.co

Total runs: 25
Run Growth: -5
Growth Rate: -19.23%
Updated:January 30 2026
huggingface.co

Total runs: 24
Run Growth: 7
Growth Rate: 30.43%
Updated:November 02 2025
huggingface.co

Total runs: 23
Run Growth: 14
Growth Rate: 60.87%
Updated:October 03 2026
huggingface.co

Total runs: 22
Run Growth: -4
Growth Rate: -16.67%
Updated:January 15 2026
huggingface.co

Total runs: 21
Run Growth: 13
Growth Rate: 61.90%
Updated:December 30 2025
huggingface.co

Total runs: 21
Run Growth: 21
Growth Rate: 100.00%
Updated:October 25 2025
huggingface.co

Total runs: 21
Run Growth: -31
Growth Rate: -140.91%
Updated:September 08 2026