Breeze TTS 2
Breeze TTS 2 is an open-weight text-to-speech model from BreezeBlue, released on 2026-08-25 with both checkpoint weights and PyTorch inference code on GitHub. It is built for real-time interaction and, at release, ranked first among open-weight systems on the Artificial Analysis TTS leaderboard while also beating several frontier proprietary systems. The single model speaks English and Chinese naturally, and it lets you shape a voice from a description or steer a cloned voice with plain language.
Key Features
- Reference-free voice design: Describe a voice in natural language and Breeze TTS 2 creates it with no reference audio at all.
- Reference-guided voice direction: Clone a voice from reference audio, then steer tone, emotion, pace, and delivery with a sentence.
- Voice clone with transcript: Pass reference audio plus its exact transcript to preserve timbre, rhythm, emotion, and style.
- Inline vocal events: Add events directly in the text, such as
(laugh),(cough),(clears throat), and(sigh)in English, or[笑],[咳嗽],[清嗓子],[叹气]in Chinese. - Ultra-low latency: Sub-40 ms time to first audio on the warmed-up fast path, and an RTF of 0.32 (roughly 3.1x real time) on an NVIDIA H100.
- GPU-efficient: Eager inference uses about 7.7 GiB of GPU memory; a 12 GB card is the minimum recommended for the eager path.
Use Cases
Who Should Use This Tool?
- Interactive voice products: Developers building assistants that need to respond without a noticeable delay.
- Game and film voice teams: Small teams that want controllable voices without hiring a voice studio.
- Localization specialists: Teams that need one model for both English and Chinese output.
Problems It Solves
- Latency kills conversation: Sub-40 ms time to first audio makes turn-taking feel natural.
- Cloning is brittle: Reference-guided direction lets you adjust a voice after cloning instead of re-recording.
- Design without a sample: Voice design gives you a starting voice from a text description alone.
Pricing
| Path | Cost | Notes |
|---|---|---|
| Weights + inference code | $0 | Weights and code are published on GitHub and Hugging Face. |
| License | Non-commercial | Inference code is Apache 2.0; model weights, derivatives, and self-hosted output are for research and non-commercial use only. |
| Compute | You pay | Runs on your own CUDA GPU; approximately 7.7 GiB for eager inference. |
Advantages & Unique Selling Points
- Ranks first among open-weight TTS on the Artificial Analysis leaderboard at release, ahead of several closed systems.
- Two control modes: Reference-free voice design and reference-guided voice direction, plus classic voice cloning.
- Genuinely bilingual: One model handles English and Chinese rather than splitting into two checkpoints.
Getting Started
- Clone breezeblue-ai/breeze-tts.
- Install the checkpoint from BreezeBlue/Breeze-TTS-2.
- Use a CUDA GPU with at least 12 GB for the eager path, or 24 GB for the fast path.
- Start with a text prompt and add inline vocal events in your language.
Frequently Asked Questions
Can I use it commercially?
The inference code is Apache 2.0, but the weights, derivatives, and self-hosted output are for research and non-commercial use only. Check the license before shipping a product.
Does it need a reference audio to create a voice?
No. Voice design works from a natural-language description; voice direction and cloning need reference audio with its transcript.
Which languages does it support?
English and Chinese with a single model.
Alternatives
- Google Gemini 3.1 Flash TTS: Commercial frontier TTS with broad language coverage.
- ElevenLabs Turbo v2.5: Proprietary voice cloning platform.
- Deepgram Nova 2: Speech-to-text, the input side of a voice pipeline.
Tips
- Use the fast path when latency matters, but budget 24 GB of GPU memory.
- Provide a transcript with your reference audio for cleaner cloning.
- Keep inline vocal events subtle; too many can sound over-produced.
Comments
No comments yet. Be the first to comment!
