yslinear/kotoba-whisper-v2.2-coreml

kotoba-whisper-v2.2 CoreML

This is the CoreML conversion of kotoba-tech/kotoba-whisper-v2.2 for use with WhisperKit.

Model Details

• Base Model: kotoba-tech/kotoba-whisper-v2.2
• Language: Japanese (ja)
• Format: CoreML (.mlmodelc)
• Optimized for: Apple Silicon (ANE - Apple Neural Engine)

Included Files

File	Description	ANE Support
AudioEncoder.mlmodelc	Audio feature encoder	100%
TextDecoder.mlmodelc	Text decoder	98%
MelSpectrogram.mlmodelc	Mel spectrogram converter	72%

Usage with WhisperKit

import WhisperKit

let whisperKit = try await WhisperKit(
    modelFolder: "path/to/kotoba-tech_kotoba-whisper-v2.2"
)

let result = try await whisperKit.transcribe(
    audioPath: "path/to/audio.wav",
    language: "ja"
)

Notes

• This is a distilled model with only 2 decoder layers (vs 32 in the original Whisper large model)
• Token-level timestamps are disabled due to alignment heads configuration incompatibility with the distilled architecture

License

This model is released under the Apache License 2.0, following the original model's license.

Attribution

This is a derivative work based on:

• kotoba-tech/kotoba-whisper-v2.2 - The original Japanese Whisper model by Kotoba Technologies
• OpenAI Whisper - The base Whisper architecture
• Distil-Whisper - Distillation codebase
• ReazonSpeech - Japanese speech dataset

Acknowledgments

• kotoba-tech for the original model
• argmaxinc for WhisperKit and whisperkittools