Higgs Audio V3 TTS

Higgs Audio V3 TTS is a conversational text-to-speech foundation model developed by Boson AI, optimized for generating naturalistic speech rather than traditional narration. Launched in June 2026, the model produces 24 kHz high-fidelity audio and supports over 100 languages. Its primary design goal is to enable voice AI agents to communicate with human-like expressiveness, incorporating emotional nuances and conversational irregularities that standard text-to-speech systems often lack.

Architecture

The model is built on a 4-billion parameter autoregressive decoder utilizing a Qwen3-4B backbone. It utilizes an interleaved tokenization scheme where text and audio tokens are processed within the same sequence. The Higgs Tokenizer encodes audio into eight discrete codebooks at 25 fps, employing a staggered delay pattern. These codes are integrated into the backbone via a fused multi-codebook embedding and head, allowing the model to ground new speech segments in both the textual input and previous audio context.

Capabilities and Control

Higgs Audio V3 features zero-shot voice cloning, enabling the replication of a target voice using only a brief reference recording. Beyond simple cloning, it offers fine-grained inline control via tags embedded directly in the text stream. These tags allow for the dynamic adjustment of:

• Emotion and Style: Adjusting the tone to be happy, sad, or adopting styles like whispering or shouting.
• Prosody and Pauses: Controlling the rhythm, emphasis, and duration of silences between words.
• Sound Effects: Triggering non-speech sounds to enhance the conversational realism.

The model is released under a research and non-commercial license, which includes a "Creator Use Grant" allowing digital content creators to use and monetize the model's output in exchange for attribution to Boson AI.