---
layout: single
title:  "FunAudioLLM：用于人类与LLM自然交互的语音理解与生成基础模型"
date:   2025-06-10 20:00:00 +0800
categories: [AI 与大模型, DevOps]
tags: [FunAudioLLM, 语音识别, 语音合成, SenseVoice, CosyVoice, FunASR, FunAudioLLM SenseVoice CosyVoice]
---

本文档介绍 [FunAudioLLM](https://funaudiollm.github.io/)，这是一个旨在增强人类与大型语言模型（LLM）之间自然语音交互的框架。其核心是两个创新模型：用于高精度多语种语音识别、情感识别和音频事件检测的 **[SenseVoice](https://github.com/FunAudioLLM/SenseVoice)**；以及用于多语种、音色和情感控制的自然语音生成的 **[CosyVoice](https://github.com/FunAudioLLM/CosyVoice)**。SenseVoice 具有极低的延迟并支持超过 50 种语言，而 CosyVoice 在多语种语音生成、零样本语音生成、跨语言语音克隆以及指令遵循能力方面表现出色。与 SenseVoice 和 CosyVoice 相关的模型已在 Modelscope 和 Huggingface 上开源，同时相应的训练、推理和微调代码也已在 GitHub 上发布。通过将这些模型与 LLM 集成，FunAudioLLM 能够实现语音翻译、情感语音聊天、交互式播客和富有表现力的有声读物叙述等应用，从而推动语音交互技术的边界。

<!-- more -->

![](/images/2025/FunAudioLLM/S2ST.png)


## SenseVoice

![](/images/2025/FunAudioLLM/SenseVoice2.png)

- [GitHub - SenseVoice](https://github.com/FunAudioLLM/SenseVoice)
- [Modelscope - SenseVoice多语言语音理解模型Small](https://www.modelscope.cn/models/iic/SenseVoiceSmall)


## [CosyVoice 2](https://funaudiollm.github.io/cosyvoice2/)

![](/images/2025/FunAudioLLM/CosyVoice2.png)

- [GitHub - CosyVoice](https://github.com/FunAudioLLM/CosyVoice)
- [Modelscope - CosyVoice语音生成大模型2.0-0.5B](https://www.modelscope.cn/models/iic/CosyVoice2-0.5B)


## 参考资料
- [FunAudioLLM: Voice Understanding and Generation Foundation Models for Natural Interaction Between Humans and LLMs](https://funaudiollm.github.io/)
- [SenseVoice](https://www.modelscope.cn/studios/iic/SenseVoice)