BreezeBlue drops Breeze-TTS-2 with 40ms real-time streaming

Breeze-TTS-2 is an open-weight text-to-speech model built specifically for real-time bilingual voice interaction.

It currently sits at the top of the Artificial Analysis open-weight TTS leaderboard, beating out several proprietary systems. The single model handles both English and Chinese with an emphasis on low latency and natural vocal delivery.

What it can do

One standout capability is prompt-driven voice design without reference audio. Instead of hunting down clean audio samples, you simply write a natural language prompt, such as describing a warm young woman with a calm delivery, and the model builds the voice from scratch.

If you do have clean reference audio, you can clone the speaker while steering delivery. Adding an instruction like speaking slowly with a serious tone directs pacing and emotion without losing the original voice identity.

It also supports inline vocal events right inside the text. Dropping parenthetical cues like (laugh), (cough), (sigh), or (clears throat) into English text, or bracketed cues like [笑] and [叹气] in Chinese, triggers realistic non-verbal sounds on cue.

Speed is the main selling point here. On a warmed-up NVIDIA H100 using its fast path, the time to first audio drops below 40 ms. It achieves a 0.32 real-time factor, generating speech at roughly 3.1 times real-time speed for continuous streaming.

Hardware and setup

Breeze-TTS-2 is relatively lean on GPU resources. In default eager mode, it pulls about 7.7 GiB of VRAM, making a 12 GB GPU the baseline requirement.

Turning on all optimization flags with static CUDA Graphs pushes VRAM usage to around 14.4 GiB. For that full fast path, you will want a 24 GB card.

Getting started is straightforward. The inference code is on GitHub and weights are hosted on Hugging Face, alongside ready-to-run Google Colab and Kaggle notebooks. BreezeBlue provides a Dockerfile configured for Hopper H100 by default, with A100 compilation flags documented.

An included single-concurrency streaming API script spins up an HTTP endpoint serving 24 kHz mono PCM audio.

Pay attention to the licensing split. The inference code uses Apache 2.0, and the audio tokenizer built on Alibaba's Qwen3-TTS is also Apache 2.0. However, the model weights and self-hosted outputs are strictly research and non-commercial. Commercial deployment requires a paid subscription to BreezeBlue's hosted API.

Why it matters

For builders testing conversational agents and companion apps, Breeze-TTS-2 cuts the feedback loop to almost nothing. You no longer need heavy external cloud APIs just to test natural, interrupted dialogue with emotional cues.

We think baking natural language voice direction and non-verbal ticks directly into a sub-40ms streaming pipeline is the right architectural move. The friction point is the license: self-hosted commercial use is locked down, so if your product scales, you will eventually have to pay for their hosted infrastructure.