Boson AI releases Higgs TTS 3, a 4B zero-shot voice cloning model for 100+ languages
Higgs TTS 3 is a roughly 4B voice-cloning TTS with inline control tags and 100+ languages; research and non-commercial licence.
Boson AI released Higgs TTS 3, described as a roughly 4B autoregressive decoder, though file metadata lists 5B. The card claims 100+ languages and reports single-digit error rates on 102 languages: 85 under 5 and 17 between 5 and 10. It offers zero-shot voice cloning, which works better with a reference transcript, and inline tags for emotion, style, prosody and sound effects. Output is 24 kHz, 25 fps with 8 codebooks and an 8,192-token context. Serving is through SGLang-Omni, vLLM-Omni or Boson AI's API.
The benchmarks are self-reported. On multilingual voice-clone error rate it scores 3.61 on Higgs-Multilingual against 3.63 for OmniVoice and 8.68 for Fish Audio S2 Pro; lower is better. The card says one error-rate table row appears to be missing a value.
The licence is the Boson Higgs TTS 3 Research and Non-Commercial License. Production use, hosted APIs and product embedding need a separate commercial licence; a Creator Use Grant allows free monetised creator content with credit. The card gives no release date; the date used is the Hugging Face repo creation date.
Strong numbers on its own card. Treat them as the maker's.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights are public with SGLang-Omni and vLLM-Omni serving; commercial use needs a deal.
- Useful 3.5: Cloning, control tags and wide language coverage fit many tasks, but only under a research licence.
- Notable 4: Over 800 likes on Hugging Face; the card reports results on 102 languages.
- Novel 3: Competent scaling of a known recipe; the language coverage stands out.
Comments