Higgs Audio v3 TTS 4B

Higgs Audio v3 TTS (4B) is a text-to-speech model that synthesizes speech from text using a Qwen3-4B backbone combined with a multi-codebook audio embedding and head, packaged as a transformers-compatible port of the original Boson AI checkpoint.

Capabilities

The model supports zero-shot TTS and voice cloning from a reference audio clip (with optional transcript). It outputs mono 24 kHz waveforms. Generation uses a delay pattern across 8 codebooks (vocabulary size 1026, including begin-of-code and end-of-code special tokens); de-delay and decoding are handled internally. The tokenizer runs at 25 frames per second — half the frame rate of many baselines — and is trained on 24 kHz data covering speech, music, and sound events in a single unified system. Its non-diffusion encoder/decoder enables fast, batch inference without iterative denoising.

Supported Languages

The model supports approximately 100 languages (87 language tags listed in the upstream metadata).

Licensing

This model is released under the Boson Higgs Audio v3 Research and Non-Commercial License. Production, hosted, or revenue-generating use requires a separate commercial license from Boson AI.

Additional Sources

The tokenizer was evaluated across DAPS (speech), MUSDB (music), and AudioSet (sound events) with 1,000 clips per category (10 seconds each), plus an Audiophile subset of 150 clips (30 seconds each) from 11 high-fidelity test discs. Metrics included acoustic reconstruction error, semantic integrity (SeedTTS subsets), and Meta Audiobox Aesthetics scores.

The model is also available via the Boson AI API (api.boson.ai/v1/audio/speech) with an OpenAI-compatible interface, preset voices, and streaming support.