← Audio

F5-TTSF5-TTS

SWivid / 上交 X-LANCE 等（开源）SWivid / SJTU X-LANCE et al. (open source)

开源MITOpen source, MIT 模型非产品Model, not a product 中文TTSChinese TTS zero-shot克隆Zero-shot cloning 自托管Self-hosted 多语言Multilingual partial lifecycle: active 核验Checked 2026-10-30

F5-TTS 是中国学术团队（上交 X-LANCE 等）的开源 flow-matching TTS 模型（MIT，★15,487，最近推送 2026-10-23 当日在册）：下载权重本机跑，中英文 zero-shot 克隆是主打用法，无官方托管服务。

F5-TTS is an open-source flow-matching TTS model from Chinese academic teams (SJTU X-LANCE et al.) (MIT, ★15,487, last push 2026-10-23 on record): download the weights and run locally, Chinese + English zero-shot cloning is the lead use, no official hosted service.

适合与不适合When to use it, when not to

适合：本机/私有化中英文 TTS 与 zero-shot 克隆研究、MIT 商用基底、大陆经 ModelScope 获取权重。不适合：要托管 API 开箱即用、要零运维、要 SLA。

Fits: local/private Chinese + English TTS and zero-shot cloning research, an MIT commercial base, mainland weight acquisition via ModelScope. Not a fit: wanting an out-of-box hosted API, zero-ops setups, SLAs.

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　中国学术团队开源项目——README 徽章当日在册 X-LANCE（上海交通大学）、SII、PCL 三实验室署名；作者 SWivid 个人维护（GitHub 主体口径）。open-source project of Chinese academic teams — README badges on record same day list X-LANCE (Shanghai Jiao Tong University), SII and PCL labs; maintained by author SWivid (GitHub entity).
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/CosyVoice 先例口径）。an open-source model is not subject to service filing (it serves no public service, same wording as the Whisper/CosyVoice precedent).

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: 开源权重线：F5-TTS（Diffusion Transformer + ConvNeXt V2）与 E2 TTS 复现两模型在册（README 当日核验）， v1 base 模型 2025-03-12 发布；下载权重本机推理，无官方托管服务。
Open-weights line: two models on record (README verified same day) — F5-TTS (Diffusion Transformer + ConvNeXt V2) and an E2 TTS reproduction; the v1 base model released 2025-03-12; download the weights and run inference locally, no official hosted service.
运行位置Where it runs: 本机 Python 推理栈（conda + PyTorch，NVIDIA CUDA 与 AMD ROCm 安装路径 README 在册，ROCm 仅 Linux）；社区 Gradio Space（E2-F5-TTS）可试跑但非官方托管产品。
Local Python inference stack (conda + PyTorch; NVIDIA CUDA and AMD ROCm install paths on record in the README, ROCm Linux-only); community Gradio Spaces (E2-F5-TTS) exist for trying it out but are not an official hosted product.
本地文件Local files: 权重、代码、生成音频全在本机；ModelScope 权重页大陆直连 200（2026-10-30）——获取路径大陆可用；无任何账户/云端数据面。
Weights, code and generated audio all stay on your machine; the ModelScope weights page measured 200 direct from mainland (2026-10-30) — the acquisition path works from the mainland; no account or cloud data surface at all.
关机后的任务Tasks after shutdown: 本机批处理推理，无任务队列形态；推理耗时随硬件变化，官方未给统一延迟口径（未核验）。
Local batch inference, no task-queue form; inference time varies with hardware and the official line gives no unified latency figure (not verified).
工具与扩展Tools & extensions: 命令行/库形态为主（README 用法示例）；zero-shot 克隆=给参考音频+文本即可合成（README 用法在册）； Sway Sampling 推理期采样策略在册。无托管 API。
CLI/library usage is the main form (usage examples in the README); zero-shot cloning = give a reference clip + text and synthesize (usage on record); the Sway Sampling inference-time sampling strategy is on record. No hosted API.
上下文与记忆Context & memory: 无会话概念——每次推理调用独立、互不携带状态；跨会话记忆不适用（本机模型形态，上下文管理完全由调用方代码自理）。
No session concept — every inference call is independent and carries no state between calls; cross-session memory does not apply (local-model form; context management is entirely up to the caller's own code).
权限与限制Permissions & limits: MIT 许可商用基底（GitHub 当日在册）；克隆授权面模型本身不做 consent 校验—— 平台侧拦截不存在（本机形态无平台），使用方自担授权责任（法律提示不在官方 README 在册，如实标未核验）。
MIT gives a commercial base (on record via GitHub same day); the model itself performs no consent check on the cloning side — there is no platform-side blocking (a local form has no platform), users carry the consent responsibility themselves (no legal notice on record in the official README, honestly marked unverified).
适合什么任务Best for: 适合：本机/私有化中英文 TTS 与 zero-shot 克隆研究、MIT 商用基底、大陆经 ModelScope 获取权重。不适合：要托管 API 开箱即用、要零运维、要 SLA。
Fits: local/private Chinese + English TTS and zero-shot cloning research, an MIT commercial base, mainland weight acquisition via ModelScope. Not a fit: wanting an out-of-box hosted API, zero-ops setups, SLAs.

头号误解Top misconception

它是模型不是产品——没有账号、控制台、SLA；「F5-TTS 服务」在官方口径里不存在（HF Space 是社区演示不是官方托管）It is a model, not a product — no account, no console, no SLA; an “F5-TTS service” does not exist in official terms (the HF Space is a community demo, not official hosting)
官方入口 github.com 大陆直连当日 000——权重经 ModelScope（AI-ModelScope/E2-F5-TTS studio 与 SWivid 权重页）直连 200，按获取路径判可达性而不是按仓库域The official entry github.com measured 000 direct from mainland on the day — weights via ModelScope (the AI-ModelScope/E2-F5-TTS studio and SWivid weight pages) measured 200 direct; judge reachability by the acquisition path, not the repo domain
README 的实验室署名（X-LANCE 上交、SII、PCL）是研究归属不是公司主体——商用集成时 MIT 无担保条款照常适用The lab signatures in the README (X-LANCE SJTU, SII, PCL) are research attribution, not a corporate entity — MIT's no-warranty terms apply as usual for commercial integration

未知项清单Known unknowns

各分发页（HF/ModelScope/Wisemodel）权重许可细节是否一致（未逐条核验）
官方对克隆使用授权的具体法律提示（README 未在册，如实标未核验）
训练数据构成与许可（论文口径未逐条核验）
Apple Silicon 路径的当前可用性（README 安装段当日只细读 NVIDIA/AMD 两节）

Whether weight-license details match across distributions (HF/ModelScope/Wisemodel, not verified item by item)
Official legal notices on cloning consent (not on record in the README, honestly marked unverified)
Training-data composition and licensing (the paper's account not verified item by item)
Current usability of the Apple Silicon path (the README install section read same day details only the NVIDIA/AMD sections)

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	F5-TTS · GitHub 仓（MIT/★15,487/最近推送 2026-10-23，GitHub API 当日核验）F5-TTS · GitHub repo (MIT / ★15,487 / last push 2026-10-23, GitHub API verified same day)	https://github.com/SWivid/F5-TTS
paper	F5-TTS · arXiv 论文（2410.06885，仓 homepage 在册）F5-TTS · arXiv paper (2410.06885, on record as the repo homepage)	https://arxiv.org/abs/2410.06885
official	F5-TTS · 官方 demo 页（swivid.github.io，2026-10-30 直连 200）F5-TTS · official demo page (swivid.github.io, direct-fetched 200 on 2026-10-30)	https://swivid.github.io/F5-TTS/
official	F5-TTS · ModelScope 权重页（2026-10-30 大陆直连 200）F5-TTS · ModelScope weights page (direct-fetched 200 from mainland on 2026-10-30)	https://www.modelscope.cn/models/SWivid/F5-TTS_Emilia-ZH-EN
changelog	F5-TTS · Releases（变更记录，2026-10-30 代理抓 200；直连 000）F5-TTS · Releases (change log, proxy-fetched 200 on 2026-10-30; direct 000)	https://github.com/SWivid/F5-TTS/releases

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）：

中英文 zero-shot 克隆的说话人相似度与稳定性（论文有基准，第三方未复核；与 CosyVoice 3.0 同场景对测最有信息量）
消费级显卡上的实时率（README 只给安装路径不给延迟口径）
v1 base 与旧 base 的音质差（2025-03-12 News 只说 better，未给对比表）

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 products).

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

Speaker similarity and stability of Chinese/English zero-shot cloning (the paper has benchmarks, no third-party re-check; head-to-head against CosyVoice 3.0 in the same scenario is the most informative)
Real-time factor on consumer GPUs (the README gives install paths but no latency figures)
Audio-quality gap between the v1 base and the older base (the 2025-03-12 news says only "better", no comparison table)