Models & agents

BreezeBlue released Breeze TTS 2, a bilingual open-weight speech model that the project says reaches 40 milliseconds to first audio on an H100

September 20, 2026 at 7:30 PM PT

GitHub social card for the breezeblue-ai/breeze-tts repository

Image: GitHub

Why it mattersA voice-agent team can now try one open-weight model that speaks two languages and follows written direction cues, without paying per second of audio to a hosted API.

BreezeBlue released Breeze TTS 2 on 2026-08-25 as a single bilingual English and Chinese text-to-speech model with Apache-2.0 inference code on GitHub at breezeblue-ai/breeze-tts. The repository has gained 465 stars in 27 days, and the README claims a rank of first among open-weight models on the Artificial Analysis text-to-speech leaderboard. That leaderboard placement has not been reproduced by an independent party yet, and the latency figures have not either, so every number below is BreezeBlue's own.

What the model does

The model handles four voice tasks from one weight set. Voice Clone copies a speaker's timbre, rhythm and delivery from a short reference clip and its transcript. Voice Design creates a voice from a written description, with no reference audio. Voice Direction clones a voice and lets a written instruction steer tone or emotion at generation time. Vocal Events lets you write markers such as "(laugh)" or "(cough)" inline in the text and hear them come out in the audio. Streaming output is 24 kHz mono PCM, which is the format a real-time phone or agent stack expects to read.

What BreezeBlue reports for speed

On an NVIDIA H100 with the warmed-up fast path enabled, BreezeBlue reports 40 millisecond time to first audio and a real-time factor of 0.32, which the project describes as 3.1 times faster than real time. Eager inference uses about 7.7 GiB of GPU memory, and the fast path uses 14.4 GiB. The listed hardware minimums are a 12 GB GPU for eager mode and a 24 GB GPU for the fast path. Linux, Python 3.10 or newer, and CUDA are required.

The licence split, written plainly

The inference code is Apache License 2.0. The model weights and checkpoints are released under the BreezeBlue Research and Non-Commercial License, meaning a team can read and audit the code freely, but cannot ship a paid product on the weights without a commercial subscription from BreezeBlue. The weights are hosted at huggingface.co/BreezeBlue/breeze-tts-2. This split is common for open-weight speech releases, and is worth reading before a team ships the model in a product that charges money.

Where the model fits

A voice-agent team building a phone bot, a live customer support layer or an in-app narrator has two options: a hosted TTS API that bills per second of audio, or a self-hosted model that runs on the team's own GPU. Breeze TTS 2 is the second option, with one weight set that speaks English and Chinese and written direction cues that change tone without recording a new voice sample. The 40 millisecond first-audio number is the figure worth measuring on the team's own hardware before shipping the model, because BreezeBlue itself is the only source for it right now.

Source

Source: BreezeBlue

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

Laya is a new open-source decision model that ConvAI says runs 7.8 times faster than Jev on the same tests

ConvAI Innovations released Laya, an Apache-2.0 decision model that outputs a probability over a fixed schema instead of text, and the repository has gained 713 stars in one day.

Source: Hacker NewsModels & agents

Nvidia says a 30B Nemotron fine-tuned on its supply-chain data scored 86.7 percent, versus 55.5 for the 550B Nemotron

Nvidia and Palantir fine-tuned a 30B Nemotron 3.5 Lightning on Nvidia's own supply-chain decisions and, according to Nvidia, it scored 86.7 percent on the internal allocation task versus 55.5 for the 550B Nemotron 3 Ultra, a model 18 times its size.

Source: PressModels & agents

Hugo Vergnes trained a 3.8B language model to 0.384 on CORE for $998 in 43 hours

Solo engineer Hugo Vergnes trained a 3.8B-parameter language model on 65B tokens in 43 hours for $998, scoring 0.384 on the CORE benchmark and beating OpenAI's 2019 GPT-2 1.5B by a wide margin.

Source: Hacker NewsModels & agents