Breeze TTS 2

Breeze TTS 2 is an open-weight, 3-billion-parameter text-to-speech model developed by BreezeBlue for real-time, expressive voice synthesis. It is designed for low-latency performance, reaching a time-to-first-audio of under 40 milliseconds. The model is bilingual, supporting English and Chinese, and has demonstrated competitive performance against both open-source and proprietary frontier systems on industry leaderboards.

The model's architecture integrates a T5-Gemma-2 text encoder with a Qwen3 backbone and a 15-step depth decoder. This structure supports three distinct control modes: Voice Clone, which replicates a speaker's timbre and style from a reference clip; Voice Design, which creates new voices from text descriptions; and Voice Direction, which allows users to steer the emotion and delivery of a voice via natural language instructions.

Additionally, Breeze TTS 2 supports inline vocal events, allowing the synthesis of non-speech sounds like laughs, sighs, and coughs by inserting parenthetical tags in English or bracketed tags in Chinese. It also supports real-time streaming at a factor of approximately 3.1x faster than playback speed. The model weights are provided for research and non-commercial use, while the accompanying inference code is released under an Apache 2.0 license.