In this categoryLocal AI · 59

More guides

Run Models

Local Apps & UIs

Hardware & Builds

Model Spotlights

Fine-Tuning

Local AIIntermediate

How to Run Breeze TTS 2 Locally: Voice Clone, Voice Design and Voice Direction

Breeze TTS 2 is a 3 billion parameter open-weight speech model that clones a voice from one reference clip and speaks English or Chinese in under 40 ms to first audio.

9 minIntermediate

Breeze TTS 2 is a 3 billion parameter open-weight text-to-speech model you can run on your own GPU. In its default eager mode it needs about 7.7 GiB of VRAM (a 12 GB card is enough); switch on the fast path and it needs 14.4 GiB (a 24 GB card). Either way it clones a voice from one short reference clip and speaks it back in under 40 ms to first audio, at a 0.32 real-time factor, in English or Chinese.

Hardware requirements

Mode	VRAM	Minimum GPU	Time to first audio
Eager	7.7 GiB	12 GB	Under 200 ms
Fast path (--fast-all)	14.4 GiB	24 GB	Under 40 ms

Where the speed claim comes from

Breeze TTS 2 ranks number one among open-weight models on the Artificial Analysis speech leaderboard, generating audio about 3.1 times faster than real time (a 0.32 real-time factor) on the fast path.

Install Breeze TTS 2

The model ships as a GitHub repo with a plain pip install, no container required. You need Python, a CUDA-capable GPU, and the driver already set up.

zsh - install

$git clone https://github.com/breezeblue-ai/breeze-tts.git

$cd breeze-tts

$python -m pip install -r requirements.txt

breezeblue-ai/breeze-ttsOfficial Breeze TTS 2 repo: install instructions, inference scripts, and the fast-path flag.github.com

Clone a voice from a reference clip

Voice cloning needs two things: a short reference audio file and the exact transcript of what is said in it. Pass both to infer.py along with the text you want spoken, and it writes a WAV file that matches the reference voice.

zsh - voice clone

$python infer.py ../breeze-tts-2 \

--ref-audio reference_en.wav \

--ref-text "This is the exact transcript of the reference clip." \

--text "Your synthesis text here" \

--output outputs/output.wav

Get the transcript exactly right

The ref-text must match the ref-audio word for word, punctuation included. A mismatched transcript is the most common cause of a cloned voice that sounds off or mispronounces words.

Add --fast-all to the command to switch on the fast path once your GPU has the 14.4 GiB of headroom it needs. It trades the extra memory for the sub-40 ms latency, which matters if you are wiring the model into a live voice agent rather than batch-rendering clips.

Design a voice from a written description

Skip the reference clip entirely and describe the voice you want in plain words. Swap --ref-audio and --ref-text for --instruction, and raise --cfg-scale to make the model follow that description more strictly.

zsh - voice design

$python infer.py ../breeze-tts-2 \

--text "(sigh) Welcome aboard, we have been expecting you." \

--instruction "A warm, thoughtful young woman with a clear voice" \

--cfg-scale 4 \

--output outputs/voice_design_en.wav

Run it as a streaming API server

For anything beyond one-off clips, start the bundled API server instead of shelling out to infer.py per request. It exposes the same voice clone and voice design paths over HTTP so other services can call it directly.

zsh - streaming server

$python -m breeze_infer.api ../breeze-tts-2 --host 0.0.0.0 --port 7860

Serving Breeze TTS 2 on http://0.0.0.0:7860

Can you use it commercially?

The inference code is Apache 2.0, so you can read it, modify it, and ship it freely. The model weights are licensed for research and non-commercial use only. To use Breeze TTS 2 in a commercial product, you need an active paid subscription through BreezeBlue's own platform rather than the free weights on Hugging Face.

Try it first without installing anything, through the official demo Space on Hugging Face.

FAQ

What GPU do I need to run Breeze TTS 2?

A 12 GB GPU covers the default eager mode at 7.7 GiB of VRAM. For the fast path's sub-40 ms latency, budget a 24 GB card to cover its 14.4 GiB footprint.

Can I use Breeze TTS 2 commercially?

Not on the free weights. The code is Apache 2.0, but the weights are licensed for research and non-commercial use, and commercial use requires a paid subscription through BreezeBlue's platform.

What languages does Breeze TTS 2 support?

English and Chinese, both from the same bilingual model, with voice cloning available in either language from a single reference clip.

Does Breeze TTS 2 run on CPU or Apple Silicon?

No. The repo requires a CUDA-capable NVIDIA GPU and Linux. There is no CPU fallback and no Apple Silicon or MPS support, so a Mac cannot run it locally today; use the Hugging Face demo Space instead.

Local models run better with more VRAM. CompareRTX GPUs on Amazonbefore you upgrade.(affiliate link. We may earn a commission at no extra cost. Disclosure)