Higgs Audio V2
Boson AI's expressive audio foundation model: a 3B Llama-3.2-based speech LLM pretrained on 10M+ hours — multi-speaker dialogue, laughter, humming, and 5-second voice cloning.
About
Higgs Audio V2 is Boson AI's bid to make expressive speech generation as open as LLMs. Instead of a narrow TTS pipeline, it's an audio foundation model: a 3-billion-parameter Llama-3.2 backbone (3.6B LLM + 2.2B audio adapter) pretrained on over 10 million hours of speech plus transcripts, lyrics, and sound-effect tags, predicting audio tokens from a 25 Hz joint acoustic-semantic codec. That LLM lineage is why it does things traditional TTS can't: clone a voice from a 5-second prompt, generate multi-speaker conversations with overlaps and laughter, follow director-style instructions ('whisper the secret, then gasp'), hum melodies in the cloned timbre, and mix speech with background music. On its own EmergentTTS-Eval harness it reports 75.7% and 55.7% win rates over gpt-4o-mini-tts on Emotions and Questions, plus strong Seed-TTS Eval and ESD numbers — all author-reported, from Boson's own eval. Licensing needs care: the code repo is Apache-2.0, but the weights carry the Boson Higgs Audio 2 Community License with a Llama-style scale gate, which several third-party writeups initially got wrong. Boson has since shipped v2.5 and a v3 (research/non-commercial), but V2 remains the widely-used open checkpoint with 8.4k GitHub stars.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- EmergentTTS-Eval (author-reported): 75.7% / 55.7% win rates over gpt-4o-mini-tts on Emotions and Questions
- Author-reported SOTA on Seed-TTS Eval and Emotional Speech Dataset (ESD)
- No independent benchmark replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: Boson AI open release (July 2025); 8.4k GitHub stars; public EmergentTTS-Eval harness; v2.5/v3 follow-ups showing active maintenance. Strengths: genuinely novel expressive capabilities (multi-speaker dialogue, paralinguistics, humming), 10M+ hour pretraining, code Apache-2.0. Docked for: (1) all benchmark numbers are author-reported with no independent replication found; (2) the license story is a known trap — weights are under the scale-gated Boson community license, NOT Apache-2.0 as early coverage claimed. License: Boson Higgs Audio 2 Community License (weights; scale-gated) — code Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Boson AI release July 2025; 3.6B LLM + 2.2B audio adapter on Llama-3.2 backbone; 10M+ hours pretraining
- · EmergentTTS-Eval (author-reported): 75.7%/55.7% win rates over gpt-4o-mini-tts on Emotions/Questions; author-reported SOTA on Seed-TTS Eval + ESD
- · Capabilities: 5-second voice cloning, multi-speaker dialogue with overlaps/laughter, director-style instructions, humming in cloned timbre
- · LICENSE TRAP: weights are under the scale-gated Boson community license, NOT Apache-2.0 as early coverage claimed — code repo is Apache-2.0
- · 8.4k GitHub stars; v2.5 and v3 (research/non-commercial) released since; V2 is the widely-used open checkpoint