Higgs Audio

Higgs Audio v2 extends Llama-3.2-3B into a text-audio foundation model, adding a 2.2B parameter DualFFN audio adapter for 5.8B parameters total and a unified tokenizer running at 25 frames per second, pretrained on 10 million hours of audio from Boson AI's AudioVerse pipeline. The result generates expressive speech with automatic prosody adaptation, zero-shot voice cloning, multi-speaker dialogue, melodic humming in a cloned voice, and speech over background music. On EmergentTTS-Eval it wins 75.71 percent of comparisons against gpt-4o-mini-tts in the Emotions category and 55.71 percent on Questions, with a 2.44 word error rate on Seed-TTS Eval. The repository installs through pip, venv, conda, or uv, includes a vLLM-backed OpenAI-compatible server for higher throughput, and recommends a GPU with at least 24 GB of memory. Code is Apache-2.0 while the weights carry the Boson Higgs Audio 2 Community License, a Llama-style grant that requires an expanded license beyond 100,000 annual active users. Boson has since shipped a standalone v3 model served through SGLang-Omni or its hosted API, so the v2 guide now lives in README_V2.md.