Zyphra releases ZONOS3-Mini, a compact open text-to-speech model
Original title: Zyphra/ZONOS3-Mini — new model trending #30 on Hugging Face
The Apache 2.0 model uses about 100 million active parameters, supports 43 languages, and can run in a WebGPU browser.
Zyphra published ZONOS3-Mini, an open-source text-to-speech model for natural speech and voice cloning from an audio sample. It has about 800 million parameters total and roughly 100 million active per token, and the card says it runs in real time on phones, laptops and in the browser. Weights are under Apache 2.0, with local browser and native runtimes plus a hosted option on Zyphra Cloud. The model supports 43 languages and word-level control of emotion, pace and delivery.
- Backbone is 11 layers with latent MoE: 64 experts and top-2 routing.
- Audio is generated at 44.1 kHz from 15 Hz dual-codec frames, up to 10 minutes.
- Voice cloning uses a reference clip up to 60 seconds or a 192-d speaker embedding.
- SGLang-Omni can serve it on NVIDIA GPUs via an OpenAI-compatible speech endpoint.
- Organizations
- Zyphra
- AI models
- ZONOS3-Mini
From the original article
4 min read at huggingface.cotext-to-speech · 4,170 downloads · 168 likes · license apache-2. 0
By Zyphra · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.