Zyphra Releases ZONOS3-Mini, a Voice-Cloning TTS Model That Runs in a Browser
Zyphra has released ZONOS3-Mini, a zero-shot voice-cloning text-to-speech model under Apache 2.0 with 800 million total and roughly 100 million active parameters, a sparse mixture of experts running 64 experts with top-2 routing. It stays real-time on consumer hardware: under 570 ms to first audio in the browser on a Pixel 9, under 150 ms on an M3 MacBook Air, and a median 80 ms on one H100 serving eight concurrent requests, where peak throughput reaches 457 seconds of audio per second. Zyphra ranks it first among open models on its own ZTTS2-Eval, ahead of Fish S2 Pro and MOSS TTS v1.5, a comparison the company publishes rather than an independent one.
A voice is cloned from a reference clip of up to 60 seconds, or built from a single 192-dimension speaker embedding, with controls for accent, mood, speaking rate and delivery down to individual words. That control rests on a new ZAC-1 codec at 15 Hz, which Zyphra says cuts word error rate 10x against DAC under the same data, and the model covers 43 languages. Runtimes ship for browsers through WebGPU, for servers through SGLang and vLLM, and for anything else through ggml; training took 112 H100s and 2.5 days. Hosted inference on Zyphra Cloud costs $10 per 1M UTF-8 bytes, capped at one concurrent request and 100 generations a day.
Related: Zyphra Releases ZAYA1-8B, a Compact MoE Model Trained on AMD, MOSS-TTS-v1.5 Tops Hugging Face Trending for Text-to-Speech