MOSS‑TTS‑Nano: Tiny Multilingual Real‑Time TTS for CPU‑Only Apps

MOSS‑TTS‑Nano is an open‑source multilingual text‑to‑speech model released by the OpenMOSS team and MOSI.AI. With only 0.1 B parameters, it targets real‑time speech generation on modest hardware – the README highlights that streaming inference runs on a 4‑core CPU without needing a GPU. The model follows a pure autoregressive pipeline that combines the MOSS‑Audio‑Tokenizer‑Nano (a lightweight 20 M‑parameter transformer‑based audio tokenizer) with a small LLM, producing 48 kHz stereo audio directly.

The model supports twenty languages, including Chinese, English, German, Spanish, French, Japanese, Korean, Russian, Arabic, and many others, as listed in the README and tags. It offers a voice‑clone mode where a reference audio clip guides the synthetic voice, and it can handle long texts via automatic chunking. Users can run the model through simple Python scripts (``infer.py`` and ``app.py``), a packaged CLI (``moss-tts-nano generate`` / ``serve``), or via the hosted Hugging Face Space demo.

MOSS‑TTS‑Nano has quickly become a trending model (over 33 k downloads and a trending score of 139) thanks to its tiny footprint, multilingual coverage, and ease of deployment for local demos, web services, and lightweight product integration. The accompanying Apache‑2.0 license and open‑source code make it suitable for both research and commercial prototypes, and the README provides clear quick‑start instructions and links to the related audio tokenizer model.