metadata
license: mit
tags:
- whisperkit
- coreml
- speech-recognition
- air-traffic-control
- atc
- aviation
language:
- en
WhisperKit ATC Models (CoreML)
CoreML-converted Whisper models fine-tuned for Air Traffic Control (ATC) transcription.
Models
| Model | Size | Recommended For | WER on ATCO2 |
|---|---|---|---|
small.en-atco2-asr |
~500MB | iPhone 12/13 | ~13.5% |
Usage with WhisperKit
import WhisperKit
let config = WhisperKitConfig(
model: "small.en-atco2-asr",
modelRepo: "skycaption/whisperkit-atc-coreml"
)
let whisperKit = try await WhisperKit(config)
Performance
These models achieve 84% better Word Error Rate compared to standard Whisper on ATC audio:
- Standard Whisper: ~50% WER on ATC
- ATC Fine-tuned: ~13.5% WER on ATCO2 dataset
Credits
- Original fine-tuning: jlvdoorn/WhisperATC (TU Delft)
- CoreML conversion: argmaxinc/whisperkittools
- App: SkyCaption
License
MIT License