← Audio

CosyVoiceCosyVoice

阿里 FunAudioLLM（开源）Alibaba FunAudioLLM (open source)

开源Open source Apache-2.0Apache-2.0 模型非产品Model, not a product 中文TTSChinese TTS 声音克隆Voice cloning 自托管Self-hosted 多语言Multilingual partial lifecycle: active 核验Checked 2026-10-30

CosyVoice 是阿里语音团队（QwenAudio/原 FunAudioLLM）的开源 LLM-TTS 模型族（Apache-2.0， v1→v2→Fun-CosyVoice 3.0 三代线）：下载权重本机推理/训练/部署全栈，中文+方言覆盖与 zero-shot 克隆是主打—— 它是模型不是产品，官方不卖托管线。

CosyVoice is the Alibaba speech team's (QwenAudio, formerly FunAudioLLM) open-source LLM-TTS model family (Apache-2.0, three generations v1 → v2 → Fun-CosyVoice 3.0): download the weights and run inference/training/deployment locally — Chinese + dialect coverage and zero-shot cloning are the pitch; it is a model, not a product, and the official line sells no hosted service.

适合与不适合When to use it, when not to

适合：本机/私有部署中文+方言 TTS、zero-shot 克隆与微调全栈、Apache-2.0 商用底座。不适合：要托管 API 开箱即用、要零运维。

Fits: local/private Chinese + dialect TTS, zero-shot cloning with a full training stack, an Apache-2.0 commercial base. Not a fit: out-of-box hosted APIs (the official line sells none), zero-ops setups.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-30

出品属地Origin: 国产domestic　阿里系语音团队开源项目（GitHub org 描述当日在册「QwenAudio Team（千问语音团队）」，关联站 funasr.com 当日直抓在册「达摩院推出的语音识别模型」叙事）。open-source project of the Alibaba speech team (GitHub org description on record same day: "QwenAudio Team" (the Chinese description adds "Qianwen speech team"); the linked site funasr.com carries the "DAMO Academy" narrative, direct-fetched 200 same day).
国内使用Direct access: 直连未核验direct unknown　官方入口 github.com 大陆直连当日实测 000（代理 200）——GitHub 大陆可达性长期波动，不下「被墙/区域限制」结论；权重经 ModelScope 分发当日直连 200，本机获取路径大陆可用。the official entry github.com measured 000 direct from mainland (200 via proxy) — GitHub's mainland reachability fluctuates long-term, no "blocked vs region-limited" conclusion drawn; weights via ModelScope measured 200 direct same day, so the local acquisition path works from the mainland.
备案状态CAC filing: 未见于备案名单not in filing list：开源模型不适用服务备案（不面向公众提供服务，同 Whisper 先例口径）；其托管形态（阿里云百炼语音合成线）的备案主语是服务方而非本开源仓。an open-source model is not subject to service filing (it serves no public service, same wording as the Whisper precedent); filing for its hosted form (Alibaba Cloud Bailian TTS line) is the service provider's, not this repo's.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 本体是开源模型族：v1（CosyVoice-300M）→ v2（CosyVoice2-0.5B）→ Fun-CosyVoice 3.0 （Fun-CosyVoice3-0.5B-2512，README 当日在册三代线）；下载权重本地推理，不接托管 API。
The core is an open-source model family: v1 (CosyVoice-300M) → v2 (CosyVoice2-0.5B) → Fun-CosyVoice 3.0 (Fun-CosyVoice3-0.5B-2512, the three-generation line on record in the README same day); download the weights and run inference locally, no hosted API.
运行位置Where it runs: 本机 Python 运行时（README 给 conda 安装+推理/训练/部署全栈，官方描述 inference, training and deployment full-stack ability）；仓已迁至 QwenAudio org（FunAudioLLM/CosyVoice 301→QwenAudio/CosyVoice，当日核验）；README 克隆命令仍写旧 org 名。
Local Python runtime (the README gives conda install + inference/training/deployment full stack, official description "inference, training and deployment full-stack ability"); the repo moved to the QwenAudio org (FunAudioLLM/CosyVoice 301 → QwenAudio/CosyVoice, verified same day); the README's clone command still writes the old org name.
本地文件Local files: 权重、代码、合成输出全在本机——开源模型仓没有账户侧数据面（这是它与托管 TTS 产品的结构性差别）。
Weights, code and synthesized output all stay on your machine — an open-source model repo has no account-side data surface (the structural difference from hosted TTS products).
关机后的任务Tasks after shutdown: 本机批处理/流式推理：README 在册 Bi-Streaming（text-in/audio-out 双流，官方称延迟低至 150ms）；无云端任务面。
Local batch/streaming inference: Bi-Streaming is on record in the README (text-in/audio-out dual streams, official claim of latency as low as 150ms); no cloud task surface.
工具与扩展Tools & extensions: 发音控制面在册（README：中文拼音/英文 CMU 音素 pronunciation inpainting、文本正则化免传统前端、 Instruct 支持情感/语速/音量指令）；MCP 不适用（模型仓不是 agent 宿主）。
Pronunciation-control surface on record (README: Chinese Pinyin / English CMU phoneme pronunciation inpainting, text normalization without a traditional frontend, Instruct support for emotion/speed/volume commands); MCP does not apply (a model repo is not an agent host).
上下文与记忆Context & memory: 不适用对话上下文维度——TTS 模型按单次合成请求工作；zero-shot 克隆以参考音频为条件（README 在册多语/跨语 zero-shot 克隆）。
The conversational-context dimension does not apply — a TTS model works per synthesis request; zero-shot cloning conditions on reference audio (multilingual/cross-lingual zero-shot cloning on record in the README).
权限与限制Permissions & limits: Apache-2.0 代码许可（GitHub API 当日核验）；声音克隆涉人格权——官方 README 未给使用授权条款，「官方未声明」按查过 README/仓根文件口径标注；9 语+18 中方言覆盖在 README 在册。
Apache-2.0 code license (GitHub API verified same day); voice cloning touches personality rights — the official README carries no usage-consent clause, labeled as "not declared by the official side" per the pages checked (README/repo root files); 9 languages + 18 Chinese dialects coverage on record in the README.
适合什么任务Best for: 适合：本机/私有部署中文+方言 TTS、要 zero-shot 克隆与训练微调全栈、要 Apache-2.0 可商用底座。不适合：要托管 API 开箱即用（官方不卖托管线）、要零运维。
Fits: local/private Chinese + dialect TTS, zero-shot cloning with a full training/fine-tuning stack, an Apache-2.0 commercial base. Not a fit: out-of-box hosted APIs (the official line sells none), zero-ops setups.

头号误解Top misconception

官方 org 已从 FunAudioLLM 更名 QwenAudio（当日 301 核验）——旧链接仍可达但新仓主在 QwenAudio；README 克隆命令仍写 FunAudioLLM 旧名，以 301 后地址为准The official org renamed from FunAudioLLM to QwenAudio (301 verified same day) — old links still resolve but the live repo is QwenAudio/CosyVoice; the README's clone command still writes the old org name, follow the post-301 address
它是模型不是产品——没有账号、没有控制台、没有 SLA；「CosyVoice 服务」在官方口径里不存在（百炼语音合成是另一条托管线，别混）It is a model, not a product — no account, no console, no SLA; a "CosyVoice service" does not exist in official terms (Bailian's TTS line is a separate hosted line, don't conflate)
官方 demo 页（funaudiollm.github.io/cosyvoice3）当日 404、新址 qwenaudio.github.io/cosyvoice3 直抓 200——README 内链仍指旧域，按新域核The official demo page (funaudiollm.github.io/cosyvoice3) 404s on the day; the new address qwenaudio.github.io/cosyvoice3 fetched 200 direct — README internal links still point to the old domain, verify against the new one

未知项清单Known unknowns

各代权重页的许可细节（仓 LICENSE=Apache-2.0 已核，ModelScope/HF 权重页未逐张核验）
平台侧克隆使用授权条款（README 未声明——按查过 README/仓根文件口径标注）
官方托管服务是否存在（百炼语音合成线是否即 CosyVoice 底座未核验到官方对应声明）

License details on each generation's weight pages (repo LICENSE=Apache-2.0 verified; ModelScope/HF weight pages not checked one by one)
Platform-side cloning usage-consent terms (not declared in the README — labeled per the pages checked)
Whether an official hosted service exists (no official statement verified mapping Bailian's TTS line to a CosyVoice base)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-30，内容更新日content updated on 2026-10-30。

类型Type	名称Name	链接Link
repo	CosyVoice · 官方仓（QwenAudio/CosyVoice，Apache-2.0，GitHub API 当日核验 ★24,192/pushed 2026-05-25）CosyVoice · official repo (QwenAudio/CosyVoice, Apache-2.0, GitHub API verified same day: ★24,192 / pushed 2026-05-25)	https://github.com/QwenAudio/CosyVoice
docs	CosyVoice · README（三代模型线/评测表/安装路径，2026-10-30 代理抓 200）CosyVoice · README (three-generation model line / eval table / install path, proxy-fetched 200 on 2026-10-30)	https://github.com/FunAudioLLM/CosyVoice
docs	Fun-CosyVoice3-0.5B-2512 · ModelScope 权重页（2026-10-30 直抓 200）Fun-CosyVoice3-0.5B-2512 · ModelScope weights page (direct-fetched 200 on 2026-10-30)	https://www.modelscope.cn/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
paper	Fun-CosyVoice 3.0 · arXiv 论文（2505.17589，2026-10-30 直抓 200）Fun-CosyVoice 3.0 · arXiv paper (2505.17589, direct-fetched 200 on 2026-10-30)	https://arxiv.org/abs/2505.17589
changelog	CosyVoice · commits（变更记录——仓无正式 releases，README Roadmap 节在册，2026-10-30 代理抓 200）CosyVoice · commits (change log — the repo has no formal releases, README Roadmap section on record; proxy-fetched 200 on 2026-10-30)	https://github.com/QwenAudio/CosyVoice/commits/main/
official	QwenAudio · 官方 demo 页（cosyvoice3，2026-10-30 直抓 200；旧域 funaudiollm.github.io 当日 404）QwenAudio · official demo page (cosyvoice3, direct-fetched 200 on 2026-10-30; old domain funaudiollm.github.io 404 same day)	https://qwenaudio.github.io/cosyvoice3/

实测记录Measured runs

本档未跑音频质量与延迟测试；结论只覆盖能力面与许可面。

实测建议（三条，都指向官方可读页没给的东西）：

官方评测表复现（README 评测表是作者自测口径——CER/SS 数字第三方未复核）
消费级 GPU 上的实时流式延迟（官方称 150ms 是特定硬件口径）
zero-shot 克隆的最低参考音频时长与质量门槛（README 未给硬指标）

This site has not completed measured runs. The coding-oriented test protocol is at protocols/coding/_protocol.md (the coding task suite does not apply to audio models).

Suggested measurements (three, all pointing at things the official readable pages do not give):

Reproducing the official evaluation table (the README's eval table is the authors' self-reported numbers — CER/SS not third-party re-checked)
Real-time streaming latency on consumer GPUs (the official 150ms claim is on specific hardware)
Minimum reference-audio length/quality for zero-shot cloning (no hard spec in the README)