iliasaz/ruaccent-coreml

RUAccentCoreML — on-device resource bundle

CoreML port of RUAccent (Russian lexical stress, variant turbo3.1) for the Swift package RUAccentCoreML. Downloaded on-device by ModelRepository.download() (swift-transformers HubApi.snapshot) and consumed by RUAccent(modelDirectory:).

Layout

• coreml/ — the four CoreML models (fp16; M2/M3 8-bit palettized): M1 accent (RoFormer), M2 omograph (BERT cross-encoder), M3 stress-usage (BERT), M4 yo-homograph (DistilBERT).
• dictpack/ — packed dictionaries (.rapack v2, mmap-able): full accents (3.19M entries), omographs, yo_words, yo_homographs.
• nn/ — the minimal tokenizer files (vocab/merges) the Swift tokenizers read.

Where this comes from, and who owns it

• Upstream model — Den4ikAI/ruaccent, MIT, © Denis Petrov. The architecture, the training and the reference implementation this port is validated against.
• Upstream weights + dictionaries — ruaccent/accentuator, MIT. The files here are those artifacts converted to CoreML and repacked; they are derivative works that inherit MIT, and copyright in them stays with the upstream authors.
• Swift runtime that loads these files — iliasaz/ruaccent-coreml (MIT).
• Downstream consumer — iliasaz/chatterbox-coreml, on-device Chatterbox TTS for Apple Silicon, which uses this for Russian stress in its multilingual pipeline.

The Swift pipeline is an independent reimplementation rather than a translation of the upstream Python. That is a description of how it was built, not a licensing necessity — upstream is MIT, so copying with attribution would have been permitted too.

Correction (2026-09-05): an earlier version of this card declared apache-2.0 and described the upstream Python as CC BY-NC-ND. Both are wrong today. Upstream relicensed to MIT in commit 39543da (2026-07-17) and ruaccent/accentuator is tagged license:mit; the CC reading came from a vendored checkout pinned at 2024-10-24.