Hlas

Hlas (Czech for "voice") is an expressive text-to-speech model published on Hugging Face by cekal, fine-tuned from BreezeBlue's Breeze TTS 2. It has 3.5B parameters, clones voices zero-shot from a short reference clip, and takes performance direction through inline tags written in square brackets. The model is released under a BreezeBlue research and non-commercial license, while its inference runtime is Apache-2.0.

What It Is

Hlas is a text-to-speech checkpoint that turns text into 24 kHz mono speech. The text carries both the words and the performance: emotions, deliveries, pauses, laughs, sighs, accents and characters are written in square brackets at the position where they should happen. The model card says it runs on a single 12 GB GPU and covers English, Mandarin and a broad set of further languages.

How Tags and Cloning Work

Tags apply from the position where they appear, can be combined, and can be free-form short descriptions. Documented families include emotion, delivery, vocalization, pacing, character and accent, and free-form directions. Tags are written in English in every language. Punctuation also shapes delivery: ellipses hesitate, capitals stress, and exclamation marks add energy.

For voice cloning, the user supplies 3 to 15 seconds of clean reference speech plus its exact transcript. The output keeps the reference voice's timbre, accent and age, and tags still apply. Without a reference, a voice can be described in text through an instruction option.

Running It

The checkpoint loads unchanged in the official Breeze TTS 2 runtime. The runtime provides a single-file inference script and a streaming server with an OpenAI-style speech endpoint that returns mono 24 kHz signed 16-bit PCM. A 12 GB GPU runs eager mode, and a 24 GB GPU can use the fast path. A live browser demo is linked from the model card.