Type a sentence and hear it spoken in the browser

Breeze TTS 2 turns text into natural speech in English and Chinese, and it gives you three ways to shape the voice. You can clone a speaker from a short reference clip and its transcript, invent a brand-new voice by describing it in plain words ("a calm older narrator with a slight rasp"), or clone a voice and then steer its tone, pace, and emotion with a written instruction. You can even drop expressive events straight into the text, like (laugh) or (sigh), and the model performs them in place.

It currently ranks first among open-weight models on the Artificial Analysis TTS leaderboard, ahead of several proprietary systems. The model is also built for real-time use: on server hardware it reaches first audio in under 40 milliseconds and generates about three times faster than playback.

Start with the demo Space: type a sentence, add a one-line voice description, and compare a few different descriptions of the same text. Then test a vocal event like (laugh) mid-sentence, since that is the feature most TTS models fail at. To run it yourself, the inference code is on GitHub and the checkpoint includes everything; you need Linux, Python 3.10+, and a CUDA GPU with roughly 8 GiB of memory free (12 GB card recommended).

The inference code is Apache 2.0, but the model weights and their outputs are licensed for research and non-commercial use only, so this is not one to ship in a paid product. Local inference needs a CUDA GPU, and the model speaks English and Chinese only.

• Narrate an article or script in a voice you designed with one sentence.
• Clone your own voice from a clean clip and read new text with it.
• Direct an existing voice: same speaker, now slower, warmer, or annoyed.
• Add laughs, sighs, and throat-clears exactly where the script needs them.
• Build a bilingual voice assistant that switches between English and Chinese.

Try the demo

View model page

Read this issue on neuralexpedition.com →