0.1B 参数跑满 20 种语言：MOSS-TTS-Nano 超轻量 TTS 开源
0.1B Parameters, 20 Languages: MOSS-TTS-Nano Ultra-Lightweight TTS Is Open Source
结论先行（BLUF）：0.1B 参数的 TTS 模型，纯 CPU 运行，输出 48kHz 双声道音频，支持 20 种语言零样本声音克隆。OpenMOSS 开源的 MOSS-TTS-Nano 不需要 GPU，不需要 PyTorch，M4 MacBook Air 单核跑起来没有压力。GitHub：OpenMOSS/MOSS-TTS-Nano
项目背景
MOSS-TTS-Nano 是 MOSI.AI 和 OpenMOSS 团队开发的语音合成模型，定位是 MOSS-TTS 系列的”微缩版”——解决 TTS 落地中的两个痛点：计算资源消耗大、部署复杂。目标是在保持实用音质的同时，做到极小资源占用 + 极低推理延迟。
核心技术架构
模型采用纯自回归（Autoregressive）架构，基于 Audio Tokenizer + LLM 流水线设计：
- 参数量：0.1B（约 1 亿），彻底脱离 GPU 依赖
- 音频输出：48kHz 采样率双声道，高保真，细节还原出色
- 推理效率：比原始版本提升近 2 倍，支持流式推理，首包速度极快
核心功能
极致轻量化 + 低延迟
- 0.1B 参数，单核 CPU（MacBook Air M4）流畅运行
- 流式推理，首包音频几乎即时生成
- ONNX Runtime CPU 运行，无需安装 PyTorch
零样本声音克隆
- Voice Clone 模式：提供一段参考音频，即可高相似度复刻声线
- 内置自动分句克隆机制，稳定处理长篇幅文本
多语言支持
支持 20 种语言：中文、英文、日语、韩语、法语、德语、西班牙语、葡萄牙语、俄语、阿拉伯语、意大利语、荷兰语、波兰语、土耳其语、越南语、泰语、印尼语、马来语、捷克语、匈牙利语
部署方式
| 方式 | 适用场景 |
|---|---|
| Python 推理脚本 | 本地快速测试 |
| FastAPI 本地 Web Demo | 可视化调试 |
| CLI 命令行工具 | 批量处理 |
| ONNX 推理方案（全套） | 无 PyTorch 生产部署 |
| MOSS-TTS-Nano-Reader | 浏览器扩展集成 |
ONNX 方案是亮点：不依赖 PyTorch 即可运行，适合嵌入浏览器扩展、本地助手、边缘计算设备。
实际意义
这个模型的价值不在于音质压倒一切——而在于让 TTS 彻底脱离 GPU 算力门槛。
几个典型应用场景：
- 本地 AI 助手语音输出：和 Ollama、LM Studio 搭配，本机闭环无需云 TTS
- 浏览器阅读插件：ONNX 直接在浏览器侧运行，无需后端服务
- 内容创作工作流：文案 → 口播音频，全链路本地
- 边缘设备：树莓派级硬件跑 TTS，传统需要云端的场景全部离线化
快速开始
git clone https://github.com/OpenMOSS/MOSS-TTS-Nano
cd MOSS-TTS-Nano
pip install -r requirements.txt
python inference.py --text "你好，世界" --language zhONNX 无 PyTorch 版本：
pip install onnxruntime
python onnx_inference.py --text "Hello world" --language en常见问题
Q：0.1B 的 TTS 音质能实用吗？
A：对于播客、旁白、工具语音提示等场景足够用。48kHz 双声道输出比大多数云端 TTS 的采样率更高。要求极高拟人度的场景（情感配音、有声书）建议搭配 7B 以上的模型。
Q：和 CosyVoice、Fish-Speech 比怎么样？
A：MOSS-TTS-Nano 的差异化是资源极限——0.1B + 纯 CPU + ONNX。CosyVoice 和 Fish-Speech 音质更优但对算力要求高。两者定位不同，不是同一赛道。
Q：ONNX 版本能在 Windows 上跑吗？
A：ONNX Runtime 跨平台，Windows/Linux/macOS 均支持。CPU-only 不挑硬件。
Q：零样本克隆的相似度如何？
A：0.1B 参数量下零样本克隆的相似度有上限，适合”风格接近”而非”完全还原”。参考音频越清晰（低噪，单人），相似度越高。
Q：能集成到 Claude Code 的语音输出流水线吗？
A：可以。通过 ONNX REST API 或 FastAPI demo 暴露 HTTP 接口，Claude Code 调用时对接即可，无需任何 GPU 资源。
© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。
BLUF: A 0.1B-parameter TTS model, CPU-only, outputting 48kHz stereo audio with 20-language zero-shot voice cloning. MOSS-TTS-Nano from OpenMOSS requires no GPU, no PyTorch, and runs comfortably on a single CPU core of an M4 MacBook Air. GitHub: OpenMOSS/MOSS-TTS-Nano
Background
MOSS-TTS-Nano is developed by MOSI.AI and OpenMOSS as the “nano” variant of the MOSS-TTS series. The goal: solve the two practical blockers for TTS deployment — high compute requirements and complex setup. Keep usable audio quality while minimizing resource footprint and inference latency.
Architecture
Pure autoregressive design based on an Audio Tokenizer + LLM pipeline:
- Parameters: 0.1B — no GPU required
- Audio output: 48kHz stereo, high-fidelity
- Inference speed: ~2× faster than the base version, with streaming support for near-instant first-chunk delivery
Key Features
Ultra-lightweight + Low Latency
- Runs on a single CPU core (tested on M4 MacBook Air)
- Streaming inference — first audio chunk is near-instant
- ONNX Runtime CPU inference — no PyTorch needed
Zero-Shot Voice Cloning
- Voice Clone mode: provide a reference audio clip → high-similarity voice replication
- Built-in sentence-splitting for stable long-text generation
Multilingual Support
20 languages: Chinese, English, Japanese, Korean, French, German, Spanish, Portuguese, Russian, Arabic, Italian, Dutch, Polish, Turkish, Vietnamese, Thai, Indonesian, Malay, Czech, Hungarian
Deployment Options
| Method | Use Case |
|---|---|
| Python inference script | Local quick test |
| FastAPI local web demo | Visual debugging |
| CLI tool | Batch processing |
| Full ONNX inference suite | PyTorch-free production deployment |
| MOSS-TTS-Nano-Reader | Browser extension integration |
The ONNX path is the standout: runs without PyTorch, suitable for browser extensions, local assistants, and edge devices.
Why It Matters
The value isn’t about competing with large TTS models on voice quality — it’s about removing the GPU compute barrier entirely.
Key use cases:
- Local AI assistant voice output: pair with Ollama or LM Studio for fully offline TTS
- Browser reading extension: ONNX runs client-side, no backend needed
- Content production pipeline: copy → audio, fully local
- Edge devices: TTS on Raspberry Pi-class hardware, formerly cloud-only scenarios moved offline
Quick Start
git clone https://github.com/OpenMOSS/MOSS-TTS-Nano
cd MOSS-TTS-Nano
pip install -r requirements.txt
python inference.py --text "Hello world" --language enONNX (no PyTorch):
pip install onnxruntime
python onnx_inference.py --text "Hello world" --language enFAQ
Q: Is 0.1B TTS audio quality actually usable?
A: Yes, for podcasts, narration, tool voice prompts, and similar use cases. 48kHz stereo output actually exceeds the sample rate of most cloud TTS APIs. For high-fidelity emotional dubbing or audiobooks, larger models (7B+) are better suited.
Q: How does it compare to CosyVoice or Fish-Speech?
A: MOSS-TTS-Nano’s differentiation is extreme resource efficiency — 0.1B + CPU-only + ONNX. CosyVoice and Fish-Speech produce higher quality audio but require significant compute. Different target use cases, not direct competitors.
Q: Does the ONNX version work on Windows?
A: Yes. ONNX Runtime is cross-platform — Windows, Linux, macOS all supported. CPU-only means hardware-agnostic.
Q: How good is zero-shot voice cloning at 0.1B?
A: There’s a ceiling at this parameter count — expect “stylistically close” rather than “exact replica.” Quality improves significantly with clean, single-speaker reference audio (low noise).
Q: Can this integrate into a Claude Code voice output pipeline?
A: Yes. Expose the ONNX inference or FastAPI demo as an HTTP endpoint, and Claude Code can call it via tool use — no GPU resources required.
© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.
💬 评论与讨论
使用 GitHub 账号登录后发表评论