ZONOS2, Zyphra's real-time voice-cloning TTS, in open weights

Zyphra releases ZONOS2, an open-weights, 8-billion-parameter MoE text-to-speech model for real-time, high-fidelity voice cloning under Apache 2.0.

Licensed under Apache 2.0, ZONOS2 is Zyphra's new real-time text-to-speech (TTS) model, focused on high-fidelity voice cloning. Its unique feature lies in its architecture: it is, according to the publisher, the first open-weights TTS built on a sparse Mixture of Experts (MoE), with 900 million active parameters out of a total of 8 billion.

This choice aims to alleviate the usual trade-off between quality and speed, with the abandonment of Classifier Free Guidance allowing, according to Zyphra, to quadruple the real-time throughput compared to the previous version. Regarding text input, the model directly reads UTF-8 bytes rather than using a phonemizer, which broadens linguistic coverage, strengthens Chinese, Korean, and Japanese support, and allows for in-sentence code-switching. It predicts Descript Audio Codec tokens for 44.1 kHz rendering and comes in two modes: one "stable," which smooths the output, and the other "expressive," which prioritizes fidelity to the clone, down to source imperfections. Trained on over six million hours of audio, ZONOS2 is accompanied by ZTTS1-Eval, a new benchmark covering up to seventeen languages and designed as a critique of existing evaluation datasets.

The model is available on Hugging Face, with its inference code on GitHub and free hosting for a limited time on Zyphra Cloud.