Guides / How to choose an open-source Chinese TTS model · 中文版

How to choose an open-source Chinese TTS model

For people who want to synthesize Chinese speech locally: picks by language, license and whether the run material is complete; no voice-quality ratings, and no promises about what has not been run.

Editorial picks

Check the card's "Languages" line first — only an explicit mention of Chinese counts; a blank means this site has not filed it, so confirm in the repo. Then the "Commercial" line: where it says check yourself, read the license before commercial use. VRAM for speech models is mostly "no estimate" and this site has not run them, so judge quality and speed by your own listening. Four picks below, each leaning a different way; open "Show how to run" to copy a command; every speech model is at /guide/tts/.

  • Instruction control, preset voices, commercial use allowed Qwen3-TTS-12Hz-1.7B-CustomVoice (Qwen)

    The material states Apache-2.0, commercial use allowed, 10 languages including Chinese, 9 preset voices, style steered by text instructions; available on ModelScope. Limit: this CustomVoice build uses preset voices — to clone your own voice, use the Base build. Community feedback: very slow inference without FlashAttention, and speech is generally too fast. This site has not run it.

  • Multilingual, conversational expressiveness; research / non-commercial license Higgs TTS 3 (Boson AI)

    The material states 100+ languages including Chinese; the license is research / non-commercial, so "Commercial" is marked check-yourself — go by the license text. No site VRAM estimate; community feedback says full precision needs 18–20GB, and an 8GB card is slow even with the quantized builds and cannot use cloning — if your GPU has under 20GB, skip it for now.

  • Small, multilingual, commercial use allowed; Chinese not listed, CPU-only limited OmniVoice (k2-fsa)

    The material states 613M parameters, Apache-2.0 commercial use allowed, 600+ languages; Chinese is not listed separately on the card — confirm in the repo examples. Installs with pip; the card has a command to copy. No site VRAM estimate; community feedback says CPU-only inference is limited and the best experience needs a GPU, with skipped or misread words on some inputs.

  • Whole-passage generation, zero-shot cloning, commercial use allowed; no streaming MOSS-TTS-v1.5 (OpenMOSS)

    The material states 8.5B, Apache-2.0 commercial use allowed, 31 languages, zero-shot cloning, available on ModelScope; the language list of the same family's Local-Transformer build names Chinese and Cantonese. No site VRAM estimate. Limit: no streaming — it generates whole passages only, latency is too high for real-time conversation, which needs the smaller MOSS-TTS-Realtime; suited to audio content.

Full comparison

  • Kokoro-82M (hexgrad)

    82M 参数轻量 TTS 模型，支持多语言，可本地部署

    Params
    82M
    Commercial
    可商用

    Run config hexgrad/Kokoro-82M

    VRAM
    no estimate
    Details
    Show how to run 
    pip install kokoro && python -c "from kokoro import KPipeline; KPipeline('a'); print('Kokoro loaded')"
    Why no estimate
    权重格式未知,无法估算显存
    License
    apache-2.0
    Languages
    8 种语言
    China access
    需代理

  • OmniVoice (k2-fsa)

    为开发者提供600+语言零样本语音合成

    Params
    613M
    Commercial
    可商用

    Run config k2-fsa/OmniVoice

    VRAM
    no estimate
    Details
    Show how to run 
    pip install omnivoice
    Why no estimate
    权重格式未知,无法估算显存
    License
    apache-2.0
    Context
    不适用
    China access
    需代理

  • Qwen3-TTS-12Hz-1.7B-CustomVoice (Qwen)

    指令式多语言TTS，流式延迟低至97ms，9音色

    Params
    1.7B
    Commercial
    可商用

    Run config Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

    VRAM
    no estimate
    Details
    Show how to run (+2 more) 
    pip install qwen-tts && python -c 'from qwen_tts import Qwen3TTSModel; model=Qwen3TTSModel.from_pretrained("Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice", device_map="cuda:0")'

    2 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache-2.0
    China access
    魔搭可用

  • Fish Audio S2 Pro (Fish Audio)

    多语种TTS，细粒度韵律/情感内联控制，80+语言

    Params
    4B + 400M
    Commercial
    需自查

    Run config fishaudio/s2-pro

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    fish-audio-research-license
    Languages
    日 / 英 / 中 (主) · 韩 / 西 / 葡 / 阿 / 俄 / 法 / 德等80+语言
    China access
    需代理

  • Higgs TTS 3 (Boson AI)

    对话式语音合成，100+语种，零样本克隆与内联控制

    Params
    4B
    Commercial
    需自查

    Run config bosonai/higgs-tts-3-4b

    VRAM
    no estimate
    Details
    Show how to run (+2 more) 
    sgl-omni serve --model-path bosonai/higgs-tts-3-4b --port 8000

    2 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    boson-higgs-audio-v3-research-and-non-commercial-license
    Context
    8192
    Languages
    100+ 语种 (含中/英/日/韩/阿拉伯等85种<5% WER)
    China access
    需代理

  • MOSS-TTS-Local-Transformer-v1.5 (OpenMOSS)

    零样本声音克隆多语言TTS模型，支持31种语言和本地部署

    Params
    4.6B
    Commercial
    可商用

    Run config OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    python -c "from transformers import AutoModel; model = AutoModel.from_pretrained('OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5', trust_remote_code=True)"

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache-2.0
    Languages
    31种语言 (含中·粤·英·日·韩·法·德·西等)
    China access
    魔搭可用

  • MOSS-TTS-v1.5 (OpenMOSS)

    零样本语音克隆多语言TTS模型，支持31种语言及显式暂停控制。

    Params
    8.5B
    Commercial
    可商用

    Run config OpenMOSS-Team/MOSS-TTS-v1.5

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    git clone https://github.com/OpenMOSS/MOSS-TTS.git && pip install -e .

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache-2.0
    China access
    魔搭可用

  • Gepard-1.0 (nineninesix)

    流式自回归 TTS，面向实时对话与语音代理

    Params
    555M
    Commercial
    需自查

    Run config nineninesix/gepard-1.0

    VRAM
    no estimate
    Details
    Show how to run 
    vllm serve nineninesix/gepard-1.0
    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache 2.0 (codec: NVIDIA Open Model License)
    Languages
    英语(美/英) · 西班牙语(墨) · 葡萄牙语(巴西) · 荷兰语
    China access
    需代理

  • Audio8-TTS-0.1B (Audio8)

    0.17B 零样本语音克隆，中英为主多语种 TTS

    Params
    0.17B
    Commercial
    需自查

    Run config Audio8/Audio8-TTS-Preview-0.1b

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Audio8 Community License v1.0
    Context
    2048
    Languages
    中文(主) · 英文(主) · 德 / 西 / 法 / 意 / 日 / 韩
    China access
    需代理

  • Dramabox (ResembleAI)

    通用文本转语音，面向配音与有声内容制作者

    Run config ResembleAI/Dramabox

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    huggingface-cli download ResembleAI/Dramabox

    1 more way(s) to run on the detail page.

    Why no estimate
    No usable VRAM basis on record
    License
    other
    Context
    不适用
    China access
    需代理

  • IndexTTS-2.5 (IndexTeam)

    零样本语音克隆，支持中英日西阿，单音频参考

    Params
    0.8B
    Commercial
    需自查

    Run config IndexTeam/IndexTTS-2.5

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    bilibili-model-license
    Languages
    中文 / 英文 / 日语 / 西班牙语 / 阿拉伯语
    China access
    魔搭可用

  • AuK-1.5B (Tencent)

    开源语音生成与编辑基础模型，支持零样本 TTS

    Params
    1.5B
    Commercial
    可商用

    Run config tencent/AuK

    VRAM
    no estimate
    Details
    Show how to run 
    python -m sglang_omni.cli serve --model-path tencent/AuK
    Why no estimate
    权重格式未知,无法估算显存
    License
    MIT
    Languages
    未公开
    China access
    魔搭可用

  • Higgs-Audio-v3-TTS-4B (Boson AI)

    语音对话TTS：100+语种、零样克隆、情感/风格内联控制

    Params
    4B
    Commercial
    需自查

    Run config bosonai/higgs-audio-v3-tts-4b

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    sgl-omni serve --model-path bosonai/higgs-audio-v3-tts-4b

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Boson Research & Non-Commercial License
    Context
    8k
    Languages
    100+语言，含中、英、日、韩、法、德等83种高质量语种
    China access
    需代理

  • Inflect-Nano-v2 (Owensong)

    4M参数本地英文TTS，CPU实时推理

    Params
    3.97M
    Commercial
    可商用

    Run config owensong/Inflect-Nano-v2

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    apache-2.0
    Context
    不适用
    Languages
    英文
    China access
    需代理

  • Audio8-TTS-Preview-0.6b (Audio8)

    0.6B多语言TTS，零样本语音克隆，支持11种语言

    Params
    0.6B
    Commercial
    可商用

    Run config Audio8/Audio8-TTS-Preview-0.6b

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache-2.0
    Context
    2048
    Languages
    粤语 / 中文 / 荷兰语 / 英语 / 法语 / 德语 / 意大利语 / 日语 / 韩语 / 波兰语 / 西班牙语
    China access
    需代理

  • scenema-audio (ScenemaAI)

    TTS 模型，面向语音合成开发者

    Run config ScenemaAI/scenema-audio

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    huggingface-cli download ScenemaAI/scenema-audio

    1 more way(s) to run on the detail page.

    Why no estimate
    No usable VRAM basis on record
    License
    other
    China access
    需代理

  • SakThai-TTS (Nanthasit)

    15语TTS GGUF，离线CPU运行

    Params
    82M
    Commercial
    需自查

    Run config Nanthasit/sakthai-tts-model

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    MIT (包装) / Apache 2.0 (Kokoro)
    Context
    不适用
    Languages
    英/日/中/韩/法/西/葡/意/德/波/俄/阿/印/孟/泰
    China access
    需代理

  • Breeze-TTS-2 (BreezeBlue)

    中英文实时TTS，支持语音克隆/设计/定向

    Params
    3.5B
    Commercial
    需自查

    Run config BreezeBlue/Breeze-TTS-2

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    研究/非商业许可；代码 Apache-2.0
    Languages
    英文 / 中文
    China access
    需代理

  • Inflect-Micro-v2 (owensong)

    9.36M 参数完整本地 TTS，固定男声英语言语合成

    Params
    9.36M
    Commercial
    可商用

    Run config owensong/Inflect-Micro-v2

    VRAM
    no estimate
    Details
    No command indexed

    No command indexed for this config. See the detail page or the official repo; commands from other configs are not borrowed.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache-2.0
    Context
    无限制（标点分块）
    Languages
    英文
    China access
    需代理

  • ZONOS2 (Zyphra)

    基于 MoE 的多语言 TTS，支持零样本克隆与流式生成

    Commercial
    可商用

    Run config Zyphra/ZONOS2

    VRAM
    no estimate
    Details
    Show how to run 
    git clone https://github.com/Zyphra/ZONOS2 && cd ZONOS2 && uv sync && uv run python -m minisgl --model-path Zyphra/ZONOS2 --tts-default-voices-dir ./default_voices/（需 NVIDIA GPU, Linux）
    Why no estimate
    No usable VRAM basis on record
    License
    apache-2.0
    Languages
    英语(主)·中文(主)·日语(主) · 韩/俄/意/葡/法/西/越/德/希伯来/荷兰/瑞典等 32 语种
    China access
    需代理

  • IndexTTS-2

    精准时长控制与情感表达的中文零样本TTS

    Commercial
    需自查

    Run config IndexTeam/IndexTTS-2

    VRAM
    no estimate
    Details
    Show how to run 
    暂无已验证的一键部署命令，需参考官方GitHub仓库（https://github.com/index-tts/index-tts）手动安装
    Why no estimate
    No usable VRAM basis on record
    China access
    魔搭可用

  • MisoTTS-8B (MisoLabs)

    基于 Sesame CSM 的对话 TTS 模型,支持语音上下文延续

    Params
    8B
    Commercial
    需自查

    Run config MisoLabs/MisoTTS

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    git clone https://github.com/MisoLabsAI/MisoTTS && cd MisoTTS && pip install -r requirements.txt

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    other
    Context
    2048
    Languages
    未声明
    China access
    需代理

  • Inflect-Nano-v1 (owensong)

    超小型英文TTS，总4.63M参数，适合本地/嵌入式实验

    Params
    4.63M
    Commercial
    可商用

    Run config owensong/Inflect-Nano-v1

    VRAM
    no estimate
    Details
    Show how to run (+1 more) 
    git clone https://huggingface.co/owensong/Inflect-Nano-v1 && cd Inflect-Nano-v1 && pip install -r requirements.txt && python inference.py --text 'Hello, world.' --out output.wav

    1 more way(s) to run on the detail page.

    Why no estimate
    权重格式未知,无法估算显存
    License
    Apache-2.0
    Languages
    英文
    China access
    需代理

Common pitfalls

  • 当前为预览版(Preview)检查点，语言覆盖和方言支持有限 — Audio8/Audio8-TTS-Preview-0.6b
  • 参考音频过长、嘈杂或转录不准会降低合成稳定性和说话人相似度 — Audio8/Audio8-TTS-Preview-0.6b
  • 需trust_remote_code=True加载，存在自定义代码安全审计需求 — Audio8/Audio8-TTS-Preview-0.6b
  • 全精度模型需 18-20GB 显存，8GB 显卡无法运行完整模式 — bosonai/higgs-tts-3-4b
  • 8GB 显卡即使使用量化版本（QT8/QT4）也较慢，需借助 CPU 内存卸载 — bosonai/higgs-tts-3-4b
  • 长文本批量生成时可能出现缺词，需用 STT 逐条校验 — bosonai/higgs-tts-3-4b
  • 8GB 显卡无法使用声音克隆功能 — bosonai/higgs-tts-3-4b
  • 仅限研究和非商业用途,生产/商用/创收需额外获取商业许可 — bosonai/higgs-audio-v3-tts-4b
  • 复杂发音(Complex Pronunciation)主观评测得分仅 25.10%,在对比模型中仍属短板 — bosonai/higgs-audio-v3-tts-4b
  • 需 48 GB 显存级别显卡才能加载,消费级 GPU 可能无法运行 — bosonai/higgs-audio-v3-tts-4b
  • 训练数据主要来自 OpenAI 和 ElevenLabs 的合成输出，存在潜在版权与许可风险 — hexgrad/Kokoro-82M
  • 对长叙事文本表现优于日常对话，对话场景下自然度可能不足 — hexgrad/Kokoro-82M
  • 需要 espeak 作为系统依赖处理 OOD 回退及部分非英语语言，增加部署复杂度 — hexgrad/Kokoro-82M
  • 最快的推理配置不一定是最佳音质，需在速度与质量之间权衡 — hexgrad/Kokoro-82M
  • 现有托管 API 服务较少，不易找到现成的云端调用方案 — hexgrad/Kokoro-82M
  • License未明确（unknown），商用需自行评估风险 — IndexTeam/IndexTTS-2
  • 显存需求较高：推荐参数（如80-200）需根据显存调整，4GB可尝试但可能受限 — IndexTeam/IndexTTS-2
  • 英文及跨语言效果论文有报但社区实测稀少，中文外场景效果不保证 — IndexTeam/IndexTTS-2
  • 无原生流式支持，完整生成耗时较长，不适合实时播报 — IndexTeam/IndexTTS-2
  • 采用 Flow Matching，不支持流式输出 — IndexTeam/IndexTTS-2
  • 音频中容易出现伪影，部分用户反馈反复调整仍无法获得自然输出，不如 Higgs Audio 效果好 — IndexTeam/IndexTTS-2
  • 官方宣传的精确音频时长控制功能实际并未支持，令用户失望 — IndexTeam/IndexTTS-2
  • 在 ComfyUI 的 TTS Audio Suite 集成中，有用户反馈情感向量节点无法影响音频输入的情感特征 — IndexTeam/IndexTTS-2
  • 部分场景下跳过或误读基础词汇，逗号后出现无法通过编辑修复的卡顿停顿 — k2-fsa/OmniVoice
  • 有用户反馈音色克隆效果“像模仿”，VibeVoice 在某些情况下更自然 — k2-fsa/OmniVoice
  • 降低 num_step 可提速但输出质量随之下降，需权衡速度与质量 — k2-fsa/OmniVoice
  • 纯 CPU 推理性能受限，最佳体验依赖 GPU 加速 — k2-fsa/OmniVoice
  • TTS 输出一致性不足,极短文本也会产生幻觉 — MisoLabs/MisoTTS
  • 有用户认为这类规模仍不够稳定,稳定性问题比想象中明显 — MisoLabs/MisoTTS
  • 以准确性换取速度，流式生成可能牺牲文本准确度 — nineninesix/gepard-1.0
  • 生产级服务化部署难度较高，涉及多阶段推理链路 — OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5
  • 生成质量不及 VibeVoice Large — OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5
  • 8B v1.5 不支持流式输出,必须一次性发送整个音频,实时语音应用里耗时太长、不够快 — OpenMOSS-Team/MOSS-TTS-v1.5
  • 实时场景只能改用更小的 MOSS-TTS-Realtime,8B 版本不适合实时生成 — OpenMOSS-Team/MOSS-TTS-v1.5
  • 训练语料生成管线、私有过滤基础设施与完整优化配方未公开,可复现性受限 — owensong/Inflect-Micro-v2
  • 固定单一音色,不可自定义语音 — owensong/Inflect-Micro-v2
  • 产品化集成前需评估局限性,不宜仅凭基准分数做决策 — owensong/Inflect-Micro-v2
  • 音质机械、有金属感，韵律平淡，不适合沉浸式听感 — owensong/Inflect-Nano-v1
  • 输出长度限制在约 15 秒 — owensong/Inflect-Nano-v1
  • 仅支持单一英语男声，无多说话人、无声音克隆、无流式输出 — owensong/Inflect-Nano-v1
  • 对长文本、罕见词、数字缩写等鲁棒性不足，独立合成片段拼接时连贯性差 — owensong/Inflect-Nano-v1
  • 目前仅支持英文（英文音素前端） — owensong/Inflect-Nano-v2
  • 当前模型音色单一，多语言和多音色尚在规划中 — owensong/Inflect-Nano-v2
  • 独立开发者项目，后续更新依赖社区反馈和兴趣 — owensong/Inflect-Nano-v2
  • 在未启用 FlashAttention 的情况下推理极慢，RTX 5090 上仅 0.3 倍实时速度，GPU 占用率仅 30% — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
  • 合成语音语速普遍偏快，无论参考语音如何，听起来都很急促 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
  • 语音克隆效果与原始音频存在可感知差异 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
  • Voice Design 模式的输出质量不适合用作训练数据 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
  • 新专区的内容全部由 AI 生成，多名用户因此表示要取消订阅、寻找替代品。 — ResembleAI/Dramabox
  • DramaBox 内容在 DramaFren 等约 18 个同类 App 中也有覆盖，独占性存疑。 — ResembleAI/Dramabox
  • 模型权重使用 LTX-2 Community License，非完全开放 — ScenemaAI/scenema-audio
  • 安装过程存在兼容性问题，依赖 nvidia-cuda-toolkit — Zyphra/ZONOS2
  • RTX 5090 等新硬件不支持 — Zyphra/ZONOS2
  • 推理速度慢 — Zyphra/ZONOS2
  • 文本开头部分可能丢失 — Zyphra/ZONOS2
  • 仅支持 Ubuntu 22.04 + NVIDIA GPU — Zyphra/ZONOS2
  • VRAM for speech models is mostly "no estimate" on this site: estimates cover only packs with a determinable weight format, and speech models often use noticeably more than a parameter-count projection. VRAM numbers in community feedback come from other people's machines — a reference, not a guarantee.
  • "Commercial: check yourself" means neither "not allowed" nor "allowed": this site has not reviewed those licenses one by one, so read the original before commercial use; this site's record may also lag the repo's current license.
  • A blank "Languages" line only means this site has not filed it; whether a model can synthesize Chinese, and how well, rests on the repo examples and your own listening — this site has done no listening comparisons.

FAQ

Can I run Chinese TTS with 4GB of VRAM?
This site has no estimate at that tier and cannot answer "yes" for you. Filter by language first, then by size: small does not mean Chinese — the Inflect models on this page, at tens of millions of parameters, state English only; rule out Higgs TTS 3, which community feedback says needs 18–20GB at full precision. Try the rest yourself, starting with rows that have a command under "Show how to run".
How do I choose a Chinese TTS model I can use commercially?
Start with the "Commercial" line: "allowed" means an Apache-2.0 or MIT-type license — Qwen3-TTS, OmniVoice and MOSS-TTS-v1.5 among the picks all are; for "check yourself", read the license. Then confirm Chinese in the "Languages" line or the repo. The repo's current license is authoritative; this site's record may lag.
No GPU at all — can it run on CPU only?
This site has no CPU-run records. Community feedback says OmniVoice is limited on CPU only and the best experience needs a GPU; models under a hundred million parameters usually demand far less hardware, but on this page their material mostly states English only. To run Chinese on CPU, look for CPU / ONNX notes in the repo first, then try it yourself.