🤖 Zyphra releases compact TTS model
ZONOS3-Mini clones voices from recordings up to 60 seconds long, sets age, accent, and timbre, and controls emotions and tempo down to individual words. Inside, it features an MoE with 64 experts and ~100 million active parameters, supporting 43 languages, including Russian.
🌍 The compact model synthesizes speech in real-time directly in the browser and on phones: on a Pixel 9, the first sound appears in 570 ms, and on a MacBook Air M3 — in 150 ms. No expensive GPU server is needed, and the open-source nature puts pressure on closed services like ElevenLabs.
👤 Available immediately: browser demo, weights on Hugging Face, and runtimes for WebGPU, MLX, and ggml. Cloning from a one-minute recording and word-level emotion tuning will be useful for dubbing, voiceovers, and assistants. Quality is confirmed only by the vendor's ZTTS2-Eval benchmark, and one should not clone someone else's voice without consent.
Source 1: https://www.zyphra.com/our-work/zonos3-mini Source 2: https://huggingface.co/Zyphra/ZONOS3-Mini