Back to Gap Map Organization

Boson AI

company · United States

Scores

1 product on the map — 1 closed.

Higgs Audio

Openness

2 medium confidence
2.0
weights
open (Higgs TTS 3 (higgs-tts-3-4b) as safetensors on the Hub, ungated)
data
closed (no training-data statement for Higgs TTS 3)
code
partial (inference through third-party serving recipes; no Higgs TTS 3 training code)
license
Boson-Higgs-TTS-3-Research-and-Non-Commercial-License (higgs-tts-3-4b, the current release of the generation line; commercial use needs a separate written license, and the free Creator Use Grant covers publishing content made with the model, not hosting it or building it into a product)
separate-line
Higgs Audio v3 STT, Apache-2.0 (four speech-to-text checkpoints (higgs-audio-v3-stt, -8b-stt and their v2 builds) released about ten weeks before Higgs TTS 3 in a collection of their own; a separate recognition line, recorded here and not scored)
superseded-release
higgs-tts-2-3b-base (Higgs Audio v2 (2025) and its tokenizer, still published under the Boson Higgs Audio 2 Community License (Llama 3 based, expanded license above 100,000 annual active users); superseded by v3 and not scored)

Higgs TTS 3, Boson's current text-to-speech release and the line its own repository now points to, is free to download but may not be used commercially without a separate license; a free grant lets creators publish and monetize content made with it. Boson's Apache-2.0 speech-to-text models are a separate line and do not change that. No training data or training code is published for Higgs TTS 3.

Adoption

3 medium confidence
3.0

Hugging Face downloads summed over the declared TTS, tokenizer and speech-to-text checkpoints; the older v2 model still draws the most.

Capability

2 medium confidence
2.0

Listeners in blind tests place Higgs TTS 3 in the lower third of the TTS arena, a step below Kokoro. Boson's separate speech-to-text model does much better on the English recognition board, but it is a different line.