CoolFace
Modelpublic

vadimsuhanov/gigaam-multilingual-large-ctc-coreml

sourceHugging Facemitupdated 2mo agoView on Hugging Face
0likes6downloads
Model Card

GigaAM Multilingual Large CTC for Core ML

Model description

This repository contains a native Core ML conversion of the multilingual GigaAM Large CTC speech-recognition model. It is intended for offline Russian and English transcription in the MeetingTranscriber macOS application.

Provenance

  • —Upstream repository: ai-sage/GigaAM-Multilingual
  • —Immutable upstream revision: 3905cd51c3ed4e88c8edf33f3302969ba480a327
  • —Upstream variant: large_ctc
  • —Conversion precision: FP32

The FP16 candidate was not published because it failed strict PyTorch/Core ML numeric and token-frame parity gates. The validated artifact uses FP32 weights.

Requirements

  • —Apple Silicon Mac
  • —macOS 26 or later
  • —Core ML ML Program runtime
  • —Mono 16 kHz PCM audio

Core ML contract

  • —Maximum input window: 25 seconds
  • —features: float32 [1, 64, 2499] log-mel features
  • —feature_length: int32 [1]
  • —logits: float32 [1, 625, 71]
  • —encoded_length: int32 [1]
  • —Vocabulary tokens: 70
  • —CTC classes: 71
  • —Blank ID: 70
  • —Greedy CTC decoding is performed by the application

Languages and Auto detection

The application exposes Russian (ru), English (en), and Auto/Bilingual. Fixed language selection controls result metadata and does not filter characters from the shared vocabulary. Auto examines recognized Unicode letters: at least 80% Cyrillic selects Russian, at least 80% Latin selects English, and mixed or empty text remains Auto.

Validation

The FP32 Core ML output was compared with the pinned PyTorch source on private Russian, English, and mixed-alphabet fixtures. The published validation report requires finite outputs, exact greedy token sequences, exact emitted token frames, bounded logit error, bounded timestamp error, and no WER regression. Fixture audio, transcripts, paths, caches, and credentials are not included.

Limitations

  • —The model accepts one batch item and at most 25 seconds per inference.
  • —Audio resampling, VAD chunking, decoding, language metadata, and word timing construction are application responsibilities.
  • —Auto detection is alphabet-based; it is not a separate acoustic language model.
  • —Validation of mixed-alphabet input does not add a third fixed language mode.

License

The upstream model and this conversion are distributed under the MIT license. See LICENSE for the complete license text.

Files

  • —model.mlpackage: FP32 Core ML ML Program
  • —model-config.json: fixed feature and encoder contract
  • —vocabulary.json: ordered 70-token vocabulary
  • —conversion-manifest.json: provenance, tensor declarations, and SHA-256 inventory
  • —validation-report.json: public, path-free parity metrics
  • —LICENSE: MIT license text