Breeze TTS 2
Breeze TTS 2 is an open-weight text-to-speech model from BreezeBlue built for real-time interaction, generating English and Chinese speech from one model with voice cloning, voice design from a text description, and inline vocal events. Its inference code is Apache-2.0, but the weights and every self-hosted output are under the BreezeBlue Research and Non-Commercial License, so commercial use needs the hosted platform or a separate agreement.
What does Breeze TTS 2 do?
Breeze TTS 2 describes itself as "an open-weight text-to-speech model built for real-time interaction." It "generates natural English and Chinese speech with a single model." It clones a voice from reference audio plus its exact transcript, designs a voice from a text description with no reference, and takes a direction instruction for tone, emotion and pace on top of a clone. Inline vocal events go in parentheses in English, such as (laugh) and (sigh), and in square brackets in Chinese.
The README's own performance figures, measured on an NVIDIA H100 with the warmed-up fast path: under 40 milliseconds to first audio and a 0.32 real-time factor. A streaming API at /v1/audio/speech returns 24 kHz 16-bit PCM at single concurrency. The README also claims the model ranks first among open-weight models on the Artificial Analysis TTS leaderboard; that is the maintainer's claim and was not checked here.
Key facts
- Licence on the inference code: Apache-2.0. The audio tokenizer is based on Qwen3-TTS, also Apache-2.0, per the README.
- Licence on the weights: the BreezeBlue Research and Non-Commercial License, read on Hugging Face; section 3 says no commercial purpose is permitted and any commercial use of the model, a derivative or an output needs a separate written licence.
- The README's own words: "The Apache License does not grant rights to use the model commercially."
- 3 billion parameters per the Hugging Face card; weights and code released 2026-08-25 per the README news.
- GPU memory about 7.7 GiB for eager inference or 14.4 GiB with --fast-all; a 12 GB GPU is the minimum recommended, 24 GB for the fast path.
- Latency figures, the project's own on an H100: under 40 ms to first audio and a 0.32 real-time factor.
- Streaming API returning 24 kHz signed 16-bit PCM at single concurrency; Docker image targets H100 by default, A100 with a build flag.
- Hosted platform: a paid subscription at breezeblue.ai grants commercial use of outputs made through the hosted API, and the README says it does not grant commercial rights to the open weights or self-hosted outputs. No tagged GitHub release; 422 stars on 2026-09-17.
What does it replace, and where does it fall short?
Breeze TTS 2 replaces ElevenLabs for research, evaluation and non-commercial work that needs low-latency bilingual English and Chinese speech with cloning and voice design on a local GPU. For a commercial product it does not, unless the audio comes through BreezeBlue's paid hosted API, which is the same shape as ElevenLabs itself.
Where it falls short: the weights are the product, and they are non-commercial, as are outputs you generate yourself. It needs a Linux machine with an NVIDIA GPU of 12 GB or more, and the streaming server handles one request at a time. Two languages. For cloning under a commercial licence, Chatterbox is the pick on the open-source alternatives to ElevenLabs page.
How does Breeze TTS 2 run?
Clone the repository, pip install -r requirements.txt, and download the checkpoint from Hugging Face. The README asks for Linux, Python 3.10 or newer, and a CUDA-capable NVIDIA GPU. A Docker image targets H100 and other Hopper GPUs by default and A100 with a build flag. infer.py is the command line; breeze_infer.api starts the single-concurrency streaming server. BreezeBlue sells a hosted platform and API whose paid subscription covers commercial use of hosted output.
Who is Breeze TTS 2 for?
A researcher or a team evaluating real-time bilingual speech on its own GPU, or a product team willing to use the hosted API for commercial output. A team that must ship self-hosted commercial audio should use Chatterbox or Kokoro instead.
What limits does the README state?
From the README: model weights, derivative models and self-hosted outputs are for research and non-commercial use only; the streaming API runs at single concurrency; the individual stage flags are for profiling and debugging; unauthorised voice cloning, impersonation and fraud are prohibited.
Questions people ask
Is Breeze TTS 2 open source?
The inference code is, under Apache-2.0. The model weights are not: they are released under the BreezeBlue Research and Non-Commercial License, and the README states that the Apache licence does not grant rights to use the model commercially. This site labels that combination open weights.
Can Breeze TTS 2 be used commercially?
Not self-hosted. The licence permits no commercial purpose for the model, derivatives or outputs without a separate written agreement. Commercial use is available through BreezeBlue's hosted platform under a paid subscription, which the README says covers hosted outputs only, not the open weights.
What hardware does Breeze TTS 2 need?
Linux, Python 3.10 or newer, and an NVIDIA CUDA GPU. The README puts memory at about 7.7 GiB for eager inference and 14.4 GiB with the fast path, and recommends a 12 GB GPU as the minimum, 24 GB for the fast path. Its latency figures were measured on an H100.
How does Breeze TTS 2 compare with ElevenLabs?
Both are commercial products at the point where audio ships: ElevenLabs through its API, Breeze through its hosted platform. The difference is that Breeze publishes the weights for research, with cloning, voice design and English plus Chinese on your own GPU. The open-source alternatives to ElevenLabs page names the projects whose weights allow shipping.
Which languages does Breeze TTS 2 speak?
English and Chinese, from one model, with inline vocal events written in parentheses for English, such as (laugh) and (sigh), and in square brackets for Chinese. The README states no other languages, and its examples cover only those two.
Sources
- Breeze TTS 2 README and Apache-2.0 LICENSE: github.com/breezeblue-ai/breeze-tts, read 2026-09-17.
- Model card and the BreezeBlue Research and Non-Commercial License: huggingface.co/BreezeBlue/breeze-tts-2, read 2026-09-17.
Compared with the others
On the open-source alternatives to ElevenLabs page, Breeze TTS 2 is the pick for real-time english and chinese, non-commercial. Open-weight bilingual model with cloning, voice design and inline vocal events; the README quotes under 40 milliseconds to first audio on an H100.
Also on that page: Chatterbox for voice cloning you can ship, Kokoro FastAPI for a speech api on your own server, Piper for small devices and offline assistants.
More text to speech and voice cloning
A multilingual text-to-speech system with voice cloning from a 10 to 30 second clip and inline emotion tags, released with its weights under a research licence that requires a separate agreement for commercial use.
An AGPL-3.0 desktop workflow engine for voice cloning, voice design, video dubbing, dictation and audiobooks, running 16 speech engines and 10 transcription engines locally, with a local API and an MCP server for agents.
A family of text-to-speech models from Resemble AI under MIT, with zero-shot voice cloning from a reference clip and a multilingual model covering 23 languages.
A research text-to-speech model with zero-shot voice cloning whose code is MIT but whose pretrained weights are non-commercial, because of the dataset they were trained on.
A fast, local neural text-to-speech engine maintained by the Open Home Foundation, used by Home Assistant and the NVDA screen reader, under GPL-3.0.
A Docker image that serves the 82-million-parameter Kokoro model through an OpenAI-compatible speech endpoint, with streaming, voice mixing and nine languages, all under Apache-2.0.
Added September 17, 2026. Every claim above comes from the project's README, LICENSE or model card, read on September 17, 2026, or from the GitHub API on the date shown in the panel. Found an error? Write to reveneau@licheo.com and it is fixed in the next weekly pass. Repository: github.com/breezeblue-ai/breeze-tts.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.