---
title: FeiyuVoice-CosyVoice3
canonical_url: "https://www.modelscope.cn/models/Esmerubphire/FeiyuVoice-CosyVoice3"
md_url: "https://www.modelscope.cn/models/Esmerubphire/FeiyuVoice-CosyVoice3.md"
repository: Esmerubphire/FeiyuVoice-CosyVoice3
last_updated: 2026-10-25
pipeline_tag: text-to-speech
tasks:
  - text-to-speech
base_model:
  - FunAudioLLM/Fun-CosyVoice3-0.5B-2512
base_model_relation: finetune
parameters: 494.0M
tensor_type:
  - BF16
library_name:
  - safetensors
  - onnx
language:
  - zh
  - en
downloads: 2407
stars: 1
tags:
  - cosyvoice3
  - character-tts
---

# FeiyuVoice-CosyVoice3

> FeiyuVoice-CosyVoice3 - Esmerubphire 在 ModelScope 开源的模型。FeiyuVoice-CosyVoice3

Esmerubphire/FeiyuVoice-CosyVoice3 是 ModelScope 魔搭社区上的 494.0M 参数text-to-speech模型，基于 FunAudioLLM/Fun-CosyVoice3-0.5B-2512 构建。

- **Repository**: Esmerubphire/FeiyuVoice-CosyVoice3
- **Tasks**: text-to-speech
- **Parameters**: 494.0M
- **Base model**: FunAudioLLM/Fun-CosyVoice3-0.5B-2512
- **Tags**: cosyvoice3, character-tts
- **Downloads**: 2407
- **Stars**: 1
- **Last updated**: 2026-10-25

Source: https://www.modelscope.cn/models/Esmerubphire/FeiyuVoice-CosyVoice3

---

# FeiyuVoice-CosyVoice3

Dual-style character TTS based on CosyVoice3.

- **Base model:** [FunAudioLLM/Fun-CosyVoice3-0.5B-2512](https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512).
- **Voices:** Cute F6 (40 updates + N01 reference); Gentle C3 (original LLM + N10 reference). Shared components are included.
- **Version:** 1.0.1 lossless. Only duplicate F6 Storage was shared again; every tensor bit is unchanged. Both voices passed strict loading and 10-condition sample-exact audio regression. No quantization or inference-math changes.

## 原始音频素材与训练授权

原始音频素材来自 B 站 UP 主 [**星落落_oi**](https://space.bilibili.com/30584746/)。**训练素材均已取得原 UP 主授权同意，用于本项目声库训练。** 感谢原作者的支持。完整署名与范围见 [COPYRIGHT.md](COPYRIGHT.md)。

## Windows + WSL（包内启动器）

Requires Windows PowerShell 5.1+, WSL2 Ubuntu 24.04, an NVIDIA GPU with a compatible WSL driver, Python 3.11, PyTorch **2.11.0+cu128**, and a separately installed [CosyVoice](https://github.com/FunAudioLLM/CosyVoice) environment. Use commit `074ca6dc9e80a2f424f1f74b48bdd7d3fea531cc` with its matching submodules and Linux dependencies. `requirements-lock.local.txt` is the validated dependency snapshot, not a portable one-command installer. Python/CUDA/source code are not bundled.

Download with the [official HF CLI](https://huggingface.co/docs/huggingface_hub/en/guides/cli):
```powershell
hf download Esmerubphire/FeiyuVoice-CosyVoice3 --local-dir FeiyuVoice-CosyVoice3
cd FeiyuVoice-CosyVoice3
Copy-Item runtime.template.json runtime.local.json
notepad runtime.local.json
```
Set `distribution`, the absolute **WSL** `python`/`repo` paths, and `python_library_dir` if your Python build needs its shared-library directory. Keep the pinned commit and Torch version. The local config is not uploaded or included in public checksums; its actual SHA is recorded in each inference result.

```powershell
.\Speak.ps1 -VerifyOnly
.\Speak.ps1 -Text '你好呀，今天想聊点什么？' -Play
.\Speak.ps1 -Style gentle -Text '慢慢来，我们还有时间。' -Play
```
Outputs are `native.wav` (24 kHz FLOAT), `speech.wav` (PCM16), and provenance JSON under `outputs/`. This is a configurable local runtime, not a standalone EXE. New-machine setup must be validated. Numeric-text behavior follows the original runtime's frontend fallback.

## 可选 CUDA／WSL ROCm 服务环境

新部署可使用配套 [声音服务与 agent 安装指南](https://github.com/Esmerubphire/astrbot_plugin_feiyu_voice/blob/main/server/INSTALL_AGENT.md)，按显卡选择 CUDA 或 WSL ROCm，并复用本仓的同一份模型权重。服务适配已提供，AMD 实体显卡尚未测试；包内原启动器仍使用上面的固定 CUDA 环境，不应通过改写包内校验文件切换后端。版本记录见 [CHANGELOG.md](CHANGELOG.md)。

## Attribution and licensing

Upstream model revision: `29e01c4e8d000f4bcd70751be16fa94bf3d85a18`. Upstream CosyVoice Apache-2.0 license and attribution are retained in `licenses/`, the base model README, and [NOTICE.md](NOTICE.md). The original audio source and confirmed training authorization are recorded in [COPYRIGHT.md](COPYRIGHT.md). No repository-wide license is assigned to local contributions or character assets; upstream Apache-2.0 remains applicable to the upstream components. The source/training attribution does not replace the existing distribution terms.