- Who it helps
- Audio developers and creators assessing self-hosted speech.
- Availability
- The model card is public. It documents a 3.5B model using BreezeTTS 2; Linux, Python 3.10+ and an NVIDIA CUDA environment are required by the runtime.
Control the performance
The card describes inline delivery tags and voice conditioning from a 3–15 second reference with an exact transcript. It prioritizes English and Mandarin and lists twenty more languages; those languages were not listening-tested here.
Separate the access terms
The runtime recommends at least 12 GB GPU memory for eager mode and 24 GB for its fastest mode. Its code license differs from its noncommercial weight and self-hosted-output terms. Commercial rights for paid hosted output do not automatically extend to local generation.
What you can do
- Read both the model card and runtime terms before choosing a deployment path. Open reference ↗
Where the evidence stops
No voice quality, timing, runtime speed or memory use was measured. The linked demo could not be opened.
Sources read. Primary documentation read October 23, 2026 UTC. No model inference or complete product workflow was executed.
Sources & further reading
Hugging Face · official
Hlas-1 model card ↗
Model description, reference-audio inputs and access requirements were read. Audio was not generated.
Vojtěch Cekal · page checked 2026-10-23
GitHub · implementation
BreezeTTS runtime and license distinctions ↗
Runtime requirements and the separate terms for code, weights and generated output were inspected.
BreezeBlue AI · page checked 2026-10-23
Coverage note: Independent demonstrations and reproduced results remain unverified; the source review does not establish a tested workflow.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Hlas-1 turns text into speech with reference-voice and delivery controls. Its local-use license matters for anyone planning to publish the output.