BreezeBlue's Breeze TTS 2 Clones and Designs Voices in Under 40ms
BreezeBlue open-sourced a 3B bilingual TTS model that tops the Artificial Analysis leaderboard with sub-40ms latency and natural-language voice design.
PRO- BreezeBlue released Breeze TTS 2, a 3B open-weight bilingual English and Chinese TTS model
- Ranks #1 on Artificial Analysis TTS leaderboard and BreezeBlue's own Voice Design, Direction, and Latency benchmarks
- Under 40 ms time to first audio and 0.32 real-time factor on a warmed-up H100
- Voice Design generates distinctive voices from natural-language prompts without reference audio
- Runs on a 12 GB GPU in eager mode using around 7.7 GiB of memory
- Weights are research and non-commercial only; commercial use requires BreezeBlue's hosted API
Text-to-speech has become one of the most competitive open-weight benchmarks in AI, and BreezeBlue just released Breeze TTS 2, a 3B-parameter speech model that combines voice cloning, natural-language voice design, and low-latency streaming in a single checkpoint runnable on a 12 GB GPU.
The model targets real-time voice AI for games, agents, and interactive stories. Voice Design generates a distinctive voice from a text prompt alone. Voice Direction then steers tone, emotion, and pace on top of any cloned identity. Both work in English and Chinese from the same checkpoint, with inline vocal events like (laugh) or (sigh) embedded directly in the text.
What's in the checkpoint
The model is a 3B-parameter transformer distributed on Hugging Face with a PyTorch inference stack. Key capabilities:
- Voice Clone from a reference audio file and its transcript
- Voice Design from a natural-language description, no reference recording needed
- Voice Direction that clones a voice while steering delivery for a specific line
- Inline Vocal Events using parentheses in English and square brackets in Chinese
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
