haykinson/whisper-large-v3-atc-coreml

Whisper Large v3 Fine-Tuned for ATC (CoreML / WhisperKit)

This repository contains Apple Silicon CoreML compiled models for jacktol/whisper-large-v3-finetuned-for-ATC, optimized for execution on Apple Neural Engine (ANE) and GPU using WhisperKit.

Model Details

• Base Model: jacktol/whisper-large-v3-finetuned-for-ATC
• Architecture: Whisper Large v3 (1.54B parameters)
• Domain: Air Traffic Control (ATC) aviation communications
• Performance: 6.5% Word Error Rate (WER) on the ATC ASR benchmark dataset

Components Included

• MelSpectrogram.mlmodelc (Audio feature extractor)
• AudioEncoder.mlmodelc (Transformer encoder on Apple Neural Engine)
• TextDecoder.mlmodelc & TextDecoderContextPrefill.mlmodelc (Text decoder)
• Tokenizer files (vocab.json, tokenizer.json, etc.)

Usage with WhisperKit (Swift)

import WhisperKit

let whisperKit = try await WhisperKit(
    modelRepo: "haykinson/whisper-large-v3-atc-coreml"
)

let results = try await whisperKit.transcribe(audioArray: pcm16kSamples)
print(results.map { $0.text }.joined())

License

MIT (Inherited from the base model).