ZONOS2

ZONOS2 is Zyphra's second-generation text-to-speech model, released June 2026, and the first open-source TTS model built on a sparse mixture-of-experts architecture: 8 billion total parameters with only 900 million active during inference. It ships two generation modes, Stable for consistent narration and Expressive for emotionally varied dialogue, and outputs 44.1 kHz audio predicted as Descript Audio Codec tokens. It supports 43 languages with tier-1 quality in English, Mandarin, and Japanese, plus zero-shot voice cloning from 10-to-30-second reference audio.

The model was trained on over 6 million hours of speech with a multi-stage cleaning pipeline, and Zyphra released a new evaluation benchmark, ZTTS1-Eval, alongside it (arXiv 2606.24320). Code is MIT and weights are Apache 2.0 on Hugging Face (Zyphra/ZONOS2), with inference code on GitHub and a free trial on Zyphra's cloud. Like all strong voice-cloning models, it carries impersonation and consent risks that deployers must manage.