openmoss/moss-tts-nano

MOSS-TTS-Nano is an open source text to speech model built by the OpenMOSS team and MOSI.AI. It is designed to be small and fast rather than to chase the largest possible model size: with only 0.1 billion parameters, it can generate speech in realtime and run directly on a CPU, so a GPU is not required. This makes it a good fit for local demos, simple web services, and lightweight product integrations where running a large model would be impractical. The model supports 20 languages, including Chinese, English, German, Spanish, French, Japanese, and several others, and it outputs 48 kHz two channel audio. Under the hood it uses an audio tokenizer paired with a language model in an autoregressive pipeline, with streaming generation that can run on a four core CPU and handle long text input through automatic chunked voice cloning. For getting started, the project ships several ways to run it: a command line script for voice cloning (infer.py), a local web demo (app.py), a packaged CLI with generate and serve commands, and an ONNX based CPU only version that removes the PyTorch dependency and roughly doubles processing efficiency, tested running smoothly on a single CPU core on a MacBook Air M4. There is also a browser extension companion project, MOSS-TTS-Nano-Reader, that runs the model directly inside the browser without a separate local server. The README documents an active release history, including finetuning code for training on custom voices, and points to demos on Hugging Face Spaces and the project's own demo site. A larger MOSS-TTS 2.0 release is described as coming soon, with the team collecting user feedback ahead of that launch. The repository is part of a small family of related MOSS-TTS and MOSS-Audio-Tokenizer models from the same team. The full README is longer than what was shown.