MacEpic/ruaccent-coreml

RUAccent for Core ML

RUAccent's Russian stress marks, exported to Core ML for TextSniper by MacEpic. TextSniper downloads it with Pocket TTS Russian, whose voice reads clearly only with stress marks, and marks the text on the Mac.

Contents

From ruaccent/accentuator (revision b78ae5ea), by Denis Petrov (Den4ikAI), licensed MIT (code) and MIT/Apache 2.0 (weights):

• homograph.mlpackage: the homograph network nn/nn_omograph/tiny2.1 (DistilBERT), fp16.
• yo.mlpackage: the ё network nn/nn_yo_homograph_resolver (DistilBERT), fp16.
• accent.mlpackage: the unknown-words network nn/nn_accent/big.onnx (RoFormer), fp32.
• The three vocabularies, homograph-vocab.txt with its two added tokens appended.
• accents.dict: the stress dictionary dictionary/accents.json.gz (3.2 million words) in a compact sorted format TextSniper reads through memory mapping.
• omographs.json, yo_words.json and yo_homographs.json: RUAccent's dictionaries, unchanged apart from being uncompressed (RUAccent's added "коса" is in omographs.json).

From CMUdict (commit 74790861), Carnegie Mellon University's pronouncing dictionary, licensed BSD-2-Clause (LICENSE-cmudict):

• cmudict.dict: the first pronunciation of each of its 124,926 words of the letters a to z, sorted by bytes. TextSniper writes the Latin words of Russian text in Cyrillic by their English pronunciation, since the Russian voice has no Latin letters (script/pocket/export_cmudict.py).

Each network was rebuilt in transformers from its config.json with its ONNX file's weights, converted with coremltools and checked against the ONNX file. bundle.json lists every file with its size and SHA-256 (18 files, 55,559,808 bytes).

License

MIT, as RUAccent: credit RUAccent and Den4ikAI. cmudict.dict is CMUdict's, under its BSD-2-Clause license in LICENSE-cmudict: Copyright (C) 1993-2015 Carnegie Mellon University. All rights reserved.