zerre-tts

zerre-tts is a Turkish text-to-speech model trained from scratch by a community author, with approximately 450 million parameters, two voice options (one male and one female), and an output sample rate of 24 kHz. The int8 ONNX version is 8.5 MB and can run in the browser using WebAssembly, as well as with ONNX Runtime or PyTorch. The training data consists of two public Turkish audiobook corpora (retaining approximately 11.9 hours), trained for about 6.5 hours on a single L4. The author reports a word error rate of 1.3% on 20 held-out segments, and 12.7% on 8 new sentences.

zerre-tts is a Turkish text-to-speech model trained from scratch by a community author, with approximately 450 million parameters, two voice options (one male and one female), and an output sample rate of 24 kHz. The int8 ONNX version is 8.5 MB and can run in the browser using WebAssembly, as well as with ONNX Runtime or PyTorch. The training data consists of two public Turkish audiobook corpora (retaining approximately 11.9 hours), trained for about 6.5 hours on a single L4. The author reports a word error rate of 1.3% on 20 held-out segments, and 12.7% on 8 new sentences.

Why it matters The author claims it is smaller than similar open-source Turkish models and is commercially usable. The architecture is adapted from tiny-voca, making it suitable for direct embedding into web pages.

Hugging Face likes 3 · Hugging Face trending score 2 · Hugging Face downloads 0

License: Commercial use allowed · ComfyUI: Not available · VRAM: Not specified

Reasons

  • Apache-2.0, commercially usable; includes an online web demo, with code and training methods publicly available on GitHub

Usage

  • Try it first in the author's web demo; locally, clone github.com/erkinemreta1/zerre-tts and then run python -m zerre.infer --checkpoint inference.pt --lang tr --text "Merhaba dünya." --voice MK --output hello.wav

Risks

  • The evaluation set is small (20 segments, 8 sentences); the author themselves notes it can only serve as a rough check. Comparisons with other models did not use the same test set.
  • Only two fixed voices are available, with no support for voice cloning or streaming. The output tends to sound like reading aloud, with limited expression for questions and exclamations. Foreign names and abbreviations are often mispronounced. Phonemization requires installing eSpeak NG separately.

Open

抱脸 text-to-speech 新模型

Explore the latest AI Radar updates