Breeze TTS 2: The Open-Weight TTS Model Topping the Leaderboard
Breeze TTS 2 is an open-weight text-to-speech model with sub-40ms latency, voice design, and a #1 spot on the Artificial Analysis leaderboard.
What is Breeze TTS 2?
Breeze TTS 2 is an open-weight text-to-speech model from BreezeBlue built for real-time, expressive voice generation. It’s ranked first among open-weight models on the Artificial Analysis TTS leaderboard, and BreezeBlue says it outperforms several proprietary frontier systems on the same benchmark. The model handles both English and Chinese, supports voice cloning from reference audio, can design entirely new voices from a text description, and streams audio with time-to-first-audio under 40 milliseconds on an H100 GPU.
TL;DR
- Breeze TTS 2 tops the open-weight rankings on the Artificial Analysis TTS Elo leaderboard and reportedly beats some closed, proprietary models on the same test.
- Latency is the headline feature, with time-to-first-audio under 40ms and a real-time factor of 0.32 (about 3.1x faster than real time) on a warmed-up H100.
- Voice design lets you skip reference audio entirely, generating a new voice purely from a natural-language description like “a warm, thoughtful young woman with a calm, reflective delivery.”
- Voice direction and voice cloning are separate tools: cloning locks in a reference speaker’s timbre and rhythm, while direction keeps that speaker’s identity but lets you steer tone, pace, and emotion with an instruction.
- Inline vocal events like
(laugh),(sigh), and(clears throat)in English, or bracketed equivalents in Chinese, can be dropped directly into the input text for expressive delivery. - Hardware requirements are modest for a real-time model: about 7.7 GiB of VRAM for eager inference (a 12GB GPU minimum) or 14.4 GiB with the full fast path enabled (24GB GPU recommended).
- The license splits code and weights: the inference code is Apache 2.0, but the model weights are restricted to research and non-commercial use, with commercial licensing available directly from the maker.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How does Breeze TTS 2 achieve such low latency?
The speed comes from a combination of architecture choices and optional CUDA-level optimizations. By default, both the command-line tool and the streaming API run in “eager” mode, meaning standard PyTorch execution without extra graph compilation. That’s the safe, portable path.
For production-grade speed, Breeze TTS 2 exposes a --fast-all flag that turns on CUDA Graph acceleration across every stage of the pipeline: the text encoder, the backbone’s prefill and decode steps, the depth decoder, and the audio codec. Each stage can also be toggled independently for profiling. When fully warmed up on an H100 with this fast path enabled, the model hits its cited numbers: under 40ms to the first audio frame and a real-time factor of 0.32, meaning it generates audio roughly 3.1 times faster than it takes to play back.
The tradeoff is memory and cold-start time. Eager inference uses about 7.7 GiB of VRAM, while the fast path pushes that to roughly 14.4 GiB, which is why BreezeBlue recommends a 24GB card if you want the full speed benefit. The warmup itself also adds startup latency, so the fast path makes sense for long-running services rather than one-off scripts.
What can you actually do with it?
Breeze TTS 2 organizes its capabilities into three distinct modes, each suited to a different use case.
Voice cloning takes a clean reference audio clip plus its exact transcript and reproduces that speaker’s timbre, rhythm, emotion, and style for new text. This is the standard TTS cloning workflow, and it requires an accurate transcript of the reference clip to work properly.
Voice design removes the reference audio requirement entirely. You describe the voice you want in plain language, matching the instruction’s language to the target text, and the model generates a voice that fits. BreezeBlue’s example instruction, “a warm, thoughtful young woman with a clear voice and a calm, reflective delivery,” produces a voice built from scratch rather than copied from a sample. A higher --cfg-scale value (the docs suggest 4) strengthens how closely the output follows the instruction.
Voice direction sits between the two. It clones a reference speaker’s identity but overlays a natural-language instruction to control delivery, so the same person’s voice can sound rushed and anxious in one generation and slow and restrained in another, without needing a differently-acted reference clip for each mood.
On top of these three modes, Breeze TTS 2 supports inline vocal events. Writing (sigh) or (clears throat) directly into English text (or the bracketed Chinese equivalents like [叹气] and [清嗓子]) inserts that non-speech sound at that point in the audio, which is a simple way to add realism without a separate emotion-control step.
How does it compare to closed TTS models?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
BreezeBlue’s core claim is that Breeze TTS 2 ranks #1 among open-weight models on the Artificial Analysis TTS leaderboard, an Elo-style ranking that pits models against each other, and that it surpasses some proprietary systems on that same benchmark. That’s a meaningful claim for a model whose weights and inference code are publicly downloadable, since real-time TTS has historically been an area where closed, API-only models led on both latency and voice quality.
BreezeBlue also published its own benchmark suite covering voice design, voice direction, and latency evaluation, which suggests the company is trying to standardize measurement for capabilities (like instruction-following in voice generation) that don’t have long-established public benchmarks the way transcription or translation do.
Because the model card frames this as a research and non-commercial release, direct side-by-side production use against commercial APIs isn’t the immediate use case. It’s more relevant to teams evaluating open alternatives, researchers benchmarking TTS instruction-following, or developers prototyping before pursuing a commercial license.
What are the licensing and hardware requirements?
The project splits its licensing cleanly down the middle. The inference code on GitHub is released under Apache 2.0, so you can freely use, modify, and redistribute the code itself. The model weights are a different matter: they’re governed by BreezeBlue’s own Research and Non-Commercial License, meaning the Apache license on the code does not extend to commercial use of the model. Anyone wanting to deploy Breeze TTS 2 in a commercial product needs written authorization from BreezeBlue’s parent company, Resonia, obtained by contacting the company directly.
On the hardware side, the model targets Linux systems with a CUDA-capable NVIDIA GPU and Python 3.10 or newer. Minimum VRAM is about 7.7 GiB for eager inference, comfortable on a 12GB card, while the fully accelerated fast path needs closer to 14.4 GiB, pointing to a 24GB card. BreezeBlue ships a Docker build targeting H100/Hopper GPUs by default, with an environment variable override for A100 support. The model also comes with responsible-use terms: users are required to secure rights and consent for any reference audio or cloned voices, and the license explicitly prohibits impersonation, fraud, and unauthorized voice cloning.
Frequently Asked Questions
Is Breeze TTS 2 free to use?
The code is Apache 2.0 licensed and free to use and modify. The model weights are free for research and non-commercial purposes only. Commercial deployment requires a paid license obtained directly from BreezeBlue’s parent company, Resonia.
Does Breeze TTS 2 support languages other than English and Chinese?
Based on the released documentation, the model is bilingual, generating natural speech in English and Chinese from a single model. No other languages are mentioned in the current release.
What GPU do I need to run Breeze TTS 2?
A 12GB GPU covers standard eager inference at about 7.7 GiB of memory use. For the accelerated CUDA Graph fast path, BreezeBlue recommends a 24GB GPU, since that mode uses roughly 14.4 GiB.
How is voice design different from voice cloning?
Voice cloning requires a reference audio clip and its exact transcript to reproduce an existing voice’s characteristics. Voice design needs no reference audio at all: you describe the voice you want in natural language and the model generates it from scratch.
How fast is Breeze TTS 2 in real-world use?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
On a warmed-up NVIDIA H100 with the fast inference path enabled, BreezeBlue reports under 40 milliseconds to first audio output and a real-time factor of 0.32, meaning the model produces audio about 3.1 times faster than the audio’s own playback duration.
