ZONOS3-Mini: Open-Source Real-Time Voice Cloning & Steerable TTS in 100M Active Parameters
Blog post from Hugging Face
Zyphra has released ZONOS3-Mini, an Apache 2.0-licensed open-source text-to-speech model for multilingual synthesis, zero-shot voice cloning, and controllable vocal delivery across 43 languages. The sparse mixture-of-experts system uses 100 million active parameters out of 800 million total, a 15 Hz dual semantic-acoustic ZAC-1 codec, and a hybrid Gated DeltaNet and attention backbone to target efficient real-time generation on devices ranging from browsers and phones to servers. Reported performance includes under 570 ms time-to-first-audio in a Pixel 9 browser, under 150 ms on an M3 MacBook Air, and 80 ms median latency with peak throughput of 457 audio seconds per second on an H100 GPU. The accompanying ZTTS2-Eval benchmark assesses speech quality, intelligibility, speaker similarity, and prosody across 55 language and acoustic subsets, where the developers report that ZONOS3-Mini ranks first among open-weight models. The release includes model weights, quantized variants, WebGPU and native runtimes, an interactive Hugging Face Space, reference code, and an expanded evaluation suite. Its controls support manipulation of speaker identity, accents, perceived age, gender presentation, vocal texture, pace, emotion, projection, and word-level delivery through embedding and activation steering, while a post-training pipeline using preference and reinforcement learning aims to improve pronunciation, robustness, and reduce hallucinations.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| Real-time | 13 | 2,990 | 636 | 190 | -28% |
| Voice AI | 7 | 1,587 | 186 | 43 | -43% |
| Vector Search | 6 | 1,097 | 201 | 87 | -45% |
| MLX | 4 | 11 | 5 | 3 | -80% |
| AI Model Fine-tuning | 1 | 613 | 130 | 56 | -29% |
| LLM | 1 | 3,355 | 742 | 191 | -49% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.