Hlas-1

Hlas-1 is an open-weight text-to-speech model with 3 billion parameters, published on Hugging Face. According to its creator, it is roughly one-tenth the size of Fish Audio's Drama-3 while remaining comparably expressive, and it accepts instructions about emotion and pacing for fine-grained control over its output.

Built as a personal side project, the model is described by its creator as imperfect but better than alternatives of similar size in the creator's experience. A demo is available at vojtechcekal.com/hlas.