Guides / How to Choose Chinese TTS Open-Source Models (2026-07) · 中文版
How to Choose Chinese TTS Open-Source Models (2026-07)
13 model cards for open-source Chinese TTS models, with key picks by use case, so you don't have to browse through each one.
Editorial picks
This page covers all 13 open-source TTS models on the site that can generate Chinese speech — 6 come with the key_info.tts field, and the other 7 (MOSS-TTS-v1.5, OmniVoice, etc.) are included through pipeline_tag because earlier card versions lacked that field. Below, picks highlight a few key ones by scenario; see the table for the rest. VRAM/language/license information is based on key_info, but when there's a conflict with community-measured results, the community measurements take precedence.
- Widest community validation / low-latency conversation Qwen3-TTS-12Hz-1.7B-CustomVoice (Qwen)
HF downloads 1.82 million, likes 1548, the largest community validation on this page; key_info.commercial_use indicates commercial use allowed, community.solves explicitly lists support for Chinese and 10 other mainstream languages, suitable for low-latency (97ms) voice interaction scenarios.
- Multilingual expressiveness / emotion control Higgs TTS 3 (Boson AI)
why_trust mentions the lowest WER across 4 benchmarks including SeedTTS/CV3, key_info.tts.languages clearly includes Chinese (100+ languages including Chinese, English, Japanese, Korean, Arabic, etc., with 85 languages having <5% WER); however, the license is boson-higgs-audio-v3-research-and-non-commercial-license, commercial_use is marked 'subject to verification', so direct commercial use is not allowed.
- Lightweight, CPU-friendly, commercial use allowed OmniVoice (k2-fsa)
key_info.size only 613M, license is apache-2.0 (commercial use allowed); community.solves mentions 'based on Qwen3-0.6B, can run on consumer hardware', watch_outs also confirms 'in environments with 8GB or less VRAM, TTS automatically offloads to CPU', making this the most hardware-friendly option on this page.
- Commercial-use long-text / multilingual cloning MOSS-TTS-v1.5 (OpenMOSS)
key_info.commercial_use indicates commercial use allowed (Apache-2.0), hf_raw.tags includes language tags like zh/yue; community.solves mentions 'provides a torch-free inference path based on llama.cpp/GGML, lowering the deployment barrier for Python environments', but watch_outs also cautions 'voice cloning quality is extremely sensitive to reference audio quality'.
Full comparison
| Model | Params | VRAM | License | Commercial use | Languages | China access | Deploy |
|---|---|---|---|---|---|---|---|
| Kokoro-82M (hexgrad) | 82M | ~0.1GB 4-bit | apache-2.0 | 需自查 | 8 种语言 | 需代理 | pip install kokoro && python -c "from kokoro import KPipeline; KPipeline('a'); print('Kokoro loaded')" |
| OmniVoice (k2-fsa) | 613M | ~0.4GB 4-bit | apache-2.0 | 可商用 | — | 需代理 | pip install omnivoice |
| Qwen3-TTS-12Hz-1.7B-CustomVoice (Qwen) | 1.7B | ~1.2GB 4-bit | Apache-2.0 | 可商用 | — | 魔搭可用 | 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")' / qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --ip 0.0.0.0 --port 8000 / modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --local_dir ./local_model |
| Higgs TTS 3 (Boson AI) | 4B | ~2.8GB 4-bit | boson-higgs-audio-v3-research-and-non-commercial-license | 需自查 | 100+ 语种 (含中/英/日/韩/阿拉伯等85种<5% WER) | 需代理 | sgl-omni serve --model-path bosonai/higgs-tts-3-4b --port 8000 / vllm-omni serve bosonai/higgs-tts-3-4b --host 0.0.0.0 --port 8095 --trust-remote-code --omni / Boson AI API (https://docs.boson.ai/models/higgs-audio-tts/overview) |
| MOSS-TTS-Local-Transformer-v1.5 (OpenMOSS) | 4.6B | ~2.7GB 4-bit | Apache-2.0 | 需自查 | 31种语言 (含中·粤·英·日·韩·法·德·西等) | 魔搭可用 | python -c "from transformers import AutoModel; model = AutoModel.from_pretrained('OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5', trust_remote_code=True)" / python -c "from modelscope import AutoModel; model = AutoModel.from_pretrained('OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5', trust_remote_code=True)" |
| MOSS-TTS-v1.5 (OpenMOSS) | 8.5B | ~5.1GB 4-bit | Apache-2.0 | 可商用 | — | 魔搭可用 | git clone https://github.com/OpenMOSS/MOSS-TTS.git && pip install -e . / transformers.AutoModel.from_pretrained('OpenMOSS-Team/MOSS-TTS-v1.5') |
| Gepard-1.0 (nineninesix) | 555M | ~0.3GB 4-bit | Apache 2.0 (codec: NVIDIA Open Model License) | 需自查 | 英语(美/英) · 西班牙语(墨) · 葡萄牙语(巴西) · 荷兰语 | 需代理 | vllm serve nineninesix/gepard-1.0 |
| Dramabox (ResembleAI) | — | — | other | — | — | 需代理 | huggingface-cli download ResembleAI/Dramabox / snapshot_download('ResembleAI/Dramabox') |
| Higgs-Audio-v3-TTS-4B (Boson AI) | 4B | ~2.8GB 4-bit | Boson Research & Non-Commercial License | 需自查 | 100+语言，含中、英、日、韩、法、德等83种高质量语种 | 需代理 | sgl-omni serve --model-path bosonai/higgs-audio-v3-tts-4b / Boson AI API (https://docs.boson.ai/models/higgs-audio-tts/overview) |
| scenema-audio (ScenemaAI) | — | — | other | — | — | 需代理 | huggingface-cli download ScenemaAI/scenema-audio / git clone https://huggingface.co/ScenemaAI/scenema-audio |
| ZONOS2 (Zyphra) | — | — | apache-2.0 | 可商用 | 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 | 需代理 | 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） |
| MisoTTS-8B (MisoLabs) | 8B | ~4.9GB 4-bit | other | 需自查 | 未声明 | 需代理 | git clone https://github.com/MisoLabsAI/MisoTTS && cd MisoTTS && pip install -r requirements.txt / huggingface-cli download MisoLabs/MisoTTS --local-dir ./weights |
| Inflect-Nano-v1 (owensong) | 4.63M | ~0.1GB 4-bit | Apache-2.0 | 需自查 | 英文 | 需代理 | 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 / git clone https://huggingface.co/owensong/Inflect-Nano-v1 && cd Inflect-Nano-v1 && pip install -r requirements.txt && python app.py |
Common pitfalls
- 仅限研究与非商业用途 — bosonai/higgs-audio-v3-tts-4b
- 部分语言(如日语)的实际表现需自行验证 — bosonai/higgs-audio-v3-tts-4b
- 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
- 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
- 语音克隆模式下无法通过文本提示调整语调，提示词仅对语音设计模式生效 — k2-fsa/OmniVoice
- 模型约 6.5GB，显存占用较大，需要与其他模型协调加载/卸载 — k2-fsa/OmniVoice
- 10 秒参考音频可能不足以捕捉情感变化，建议 30-60 秒以上以获得更可靠的音色、节奏和说话风格 — k2-fsa/OmniVoice
- 要获得最佳且最一致的克隆效果，需要用 45 分钟以上、包含丰富情感变化的音频进行微调 — k2-fsa/OmniVoice
- 8GB 或更低显存环境下，TTS 会自动卸载到 CPU 运行 — k2-fsa/OmniVoice
- 官方说明语音克隆是最稳定的模式，暗示其他模式可能稳定性不足 — k2-fsa/OmniVoice
- 短文本也会出现幻觉/乱读，需通过减少大写等方式缓解但无法根除 — MisoLabs/MisoTTS
- 合成一致性不足，同一段话容易出现偏离 — MisoLabs/MisoTTS
- 作为 8B 参数模型，实际表现让人觉得训练不充分 — MisoLabs/MisoTTS
- 参考 Python 推理代码速度仅 RTF ~2（约 50% 实时），本地实时使用有压力 — OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5
- 服务化部署复杂，需处理参考音频编码、自回归生成、多码本采样与有状态解码器等多阶段流水线 — OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5
- 不指定语言标签时，部分语言质量可能相比 1.0 版本回退 — OpenMOSS-Team/MOSS-TTS-v1.5
- 生成质量对部署配置敏感，需调优才能达到最佳效果 — OpenMOSS-Team/MOSS-TTS-v1.5
- 音质存在可感知的缺陷,非生产级/非 SOTA — owensong/Inflect-Nano-v1
- 目前仅为实验性发布,作者表示若有足够关注才考虑加大训练预算做 v2 — owensong/Inflect-Nano-v1
- 推理速度偏慢，有用户反馈生成速度低 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
- GPU 利用率仅约 12%，存在优化空间 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
- 与 VibeVoice 相比表达力偏平淡/单调 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
- VoiceDesign 模型设计的语音难以在下一次生成中保留复用 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
- 社区反馈缺少阿拉伯语、越南语、荷兰语、罗马尼亚语、粤语等语言支持 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
- 社区 faster 版本在 PyTorch ≤2.5.0 上 CUDA graph 捕获不可靠 — Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
- 新栏目内容被指为 AI 生成，非真人出演 — ResembleAI/Dramabox
- 部分用户因此取消订阅并寻求替代应用 — ResembleAI/Dramabox
- 原生克隆(native cloning)方案仍在探索替代路线,当前方案可能并非最终形态 — ScenemaAI/scenema-audio
- 社区关注度极低,缺乏独立验证与实测反馈 — ScenemaAI/scenema-audio
- 免费额度仅 100 分钟/月，重度使用不够 — Zyphra/ZONOS2
- 生成偶尔不稳定，可能出现错误（官方自述 'sometimes unreliable in generations'） — Zyphra/ZONOS2
- The VRAM numbers in the table are estimates from a parameter quantization formula; actual TTS/speech model usage may be significantly higher — for example, k2-fsa/OmniVoice's structured field states Q4 ~0.4GB, but community measurements show real usage around 6.5GB and it automatically switches to CPU below 8GB, so please rely on community measurements or your own tests rather than just this column's numbers.
FAQ
- Can I run Chinese TTS with 4GB VRAM?
- Yes, but you need to pick the right models. owensong/Inflect-Nano-v1 (4.63M parameters), Aratako/Irodori-TTS-500M-v3 (key_info.vram about Q4 0.3GB) are very small; k2-fsa/OmniVoice's structured field says Q4 ~0.4GB, but community tests show real usage around 6.5GB and it automatically switches to CPU when below 8GB — in such conflicts, follow community measurements; 4GB VRAM is better suited to those definitively smaller models.
- How to choose a Chinese TTS model that allows commercial use?
- In the table, key_info.commercial_use marked 'commercial use allowed' includes k2-fsa/OmniVoice, Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice, OpenMOSS-Team/MOSS-TTS-v1.5, Zyphra/ZONOS2, Aratako/Irodori-TTS-500M-v3 (Japanese specialization); Boson AI's Higgs series and MOSS-TTS-Local-Transformer-v1.5 are marked 'subject to verification', so please check the license terms yourself before commercial use.
- Can I run it on CPU only, without any GPU?
- Supertone/supertonic-3 is an ONNX model; community.solves confirms it can run offline without GPU (99M parameters); k2-fsa/OmniVoice will also automatically fall back to CPU when VRAM is below 8GB, but speed will decrease.