Home/Voice and audio/Hlas-1
Hlas-1
A 3.5 billion parameter text-to-speech model with zero-shot voice cloning. The license is non-commercial.
Posted on Show HN on October 23, 2026. The Hugging Face repo was created on October 19. Hlas-1 is a 3.5 billion parameter expressive text-to-speech model. It clones a voice from a 3 to 15 second reference clip and accepts inline tags such as [whispers], [laughs], and [long pause]. English and Mandarin come first, plus many other languages. Output is 24 kHz. The model card says it runs on a single 12 GB GPU and uses the Breeze TTS 2 runtime. The license is BreezeBlue Research and Non-Commercial, so commercial use is not allowed. A demo is at vojtechcekal.com/hlas.
Where it ranks
08 in Voice and audio
Expressive speech and a cloned voice from a short clip, for non-commercial use.
See the voice and audio list
Pricing
Check site
The license is BreezeBlue Research and Non-Commercial. Commercial use is not allowed. Check the model card.
Checked
Added October 23, 2026
Last verified October 23, 2026
Dated releases
Hlas-1 was posted as a 3.5 billion parameter text-to-speech model with zero-shot voice cloning, under a non-commercial license.
Also in voice and audio
- ElevenLabs · Text to speech, voice tools, and an API for putting speech in a product.
- Suno · Makes a full song, vocals included, from a description or your lyrics.
- Descript · Edit video and podcasts by editing the transcript text.