tiny-voca-ru / NOTICE
egnch's picture
Add INT8 tiny2.1 onnx/omograph.onnx and update docs for stress=full (v0.2.0)
be075ba verified
Raw History Blame Contribute Delete
1.17 kB
tiny-voca-ru
A lightweight 4.5M-parameter Russian text-to-speech model.
This project is based on tiny-voca:
tiny-voca, Copyright 2026 Muhammed Nazeem
https://github.com/n4ze3m/tiny-voca
Apache License 2.0, commit db851f75c6287a15be4c4e995f931bce3a86b47b
The model architecture and training recipe originate from tiny-voca.
Changes for Russian include the pure-Python Russian text frontend
(src/tiny_voca_ru/frontend/: text normalization, stress placement, G2P rules,
and 61-symbol Russian phoneme inventory), Russian forced-alignment calibration,
safetensors checkpoint support, ONNX inference/export, and the Russian browser
demo.
The code is licensed under the Apache License 2.0 (see LICENSE).
The stress dictionary (stress.txt.gz), unknown-word accent model (accent.onnx),
homograph tokenizer vocabulary (omograph_vocab.txt.gz), and INT8-quantized
homograph resolver (omograph.onnx, derived from nn/nn_omograph/tiny2.1) are
derived from ruaccent:
https://github.com/Den4ikAI/ruaccent
https://huggingface.co/ruaccent/accentuator
MIT License (see src/tiny_voca_ru/frontend/data/LICENSE-ruaccent).
tiny-voca-ru is not an official tiny-voca release.