Hlas-1
Expressive, controllable speech in a 3.5B open-weight model
Start new thread

Hlas-1 - Expressive, controllable speech in a 3.5B open-weight model

by•
Hlas-1 is a 3.5B open-weight text-to-speech model that whispers, laughs, cries, and follows custom emotion and pacing instructions. Shape delivery with inline tags, clone a voice from a short reference clip, and generate multilingual speech. Built as a spare-time project on Breeze TTS 2, with 24 kHz audio and a streaming runtime. Free for research and non-commercial use.

Add a comment

Replies

Best
Maker
📌
Hey Product Hunt, I’m Vojtech 👋 $0 seed round. 0 users. Just a fun side project built in my spare time. Meet Hlas-1: a 3.5B open-weight text-to-speech model. Hlas means “voice” in Czech. I fine-tuned it from Breeze TTS 2 to make speech expressive and controllable. The words and the performance live in the same text. Add [whispers], [laughs], [crying], or [long pause] exactly where you want them. You can also write your own directions, like [nervous, trying to sound confident] or [starting calm, then losing patience]. It also supports voice cloning from a short reference clip, multilingual speech, and 24 kHz audio through a streaming runtime. Try it: Weights: Hlas-1 is derived from Breeze TTS 2 by BreezeBlue and licensed for research and non-commercial use only. The inference runtime is Apache-2.0. I’d love to hear what you make with it, especially unusual emotion or pacing directions. Which line should I try next?