Boson AI's Higgs Audio v3 Speaks 15x Better Across 100 Languages
Boson AI's 4B open-weight TTS model speaks before a sentence finishes, clones voices in 100 languages, and runs faster than real-time on a single H100
Higgs Audio v3 TTS is Boson AI's new open-weight text-to-speech model, and it was built around one core insight: voice agents don't need a model that reads text aloud, they need one that actually speaks. That means starting audio before a sentence is complete, reacting with the right emotion, and keeping up with a live conversation without sounding robotic.
Higgs Audio v3 is built for voice chat: it speaks, not just reads, turning model responses into expressive conversational speech across 100 languages, with zero-shot voice cloning and inline control over emotion, style, prosody, pauses, and sound effects.
A model that doesn't wait for a full sentence
Higgs Audio v3 was designed for streaming interaction. It can begin synthesis before a full sentence or punctuation mark arrives, then continue as the text stream grows while preserving a stable delivery. This is a fundamental shift from traditional TTS pipelines that buffer a complete utterance before generating any audio.
Voice AI needs a different kind of text-to-speech. In a live conversation, speech is not just the last step after text generation. It is how the agent answers, reacts, pauses, emphasizes, and carries the turn.
Under the hood: a Qwen3 backbone with audio codebooks
Architecturally, Higgs is a roughly 4B-parameter autoregressive decoder built on a Qwen3-4B backbone. It consumes interleaved text and audio tokens. Audio is encoded by the Higgs Tokenizer into 8 discrete codebooks at 25 fps, staggered with a delayed pattern, mapped into the backbone hidden states through a fused multi-codebook embedding, and decoded back into a 24 kHz waveform through a fused multi-codebook head.
Generation alternates between text and audio chunks, so each new audio segment is grounded in both the reference audio and the context generated so far. The delay pattern across codebooks is the key trick here: it lets the model generate multiple audio streams in parallel without losing coherence, enabling the low-latency streaming behavior.