mustafademir/whisperkit-coreml

WhisperKit Core ML models

Core ML (WhisperKit format) conversions of Whisper models, for on-device speech recognition on Apple platforms.

Variants

selimc_whisper-large-v3-turbo-turkish

Core ML conversion of selimc/whisper-large-v3-turbo-turkish by Selim Çavaş — a Turkish fine-tune of openai/whisper-large-v3-turbo, trained on Common Voice 17.0 (tr).

• Precision: float16
• Files: AudioEncoder.mlmodelc, TextDecoder.mlmodelc, MelSpectrogram.mlmodelc, TextDecoderContextPrefill.mlmodelc
• Size: ~1.5 GB
• Converted with whisperkittools (whisperkit-generate-model)

Conversion accuracy vs. the PyTorch reference (torch2coreml PSNR): TextDecoder 38.5 dB, AudioEncoder 52.8 dB, MelSpectrogram 67.8 dB, decoder context prefill caches 88.6 / 92.5 dB.

Usage

let folder = try await WhisperKit.download(
    variant: "selimc_whisper-large-v3-turbo-turkish",
    from: "mustafademir/whisperkit-coreml"
)
let pipe = try await WhisperKit(WhisperKitConfig(modelFolder: folder.path, load: true, download: false))

License

MIT, inherited from the base model. All credit for the fine-tune goes to Selim Çavaş; this repository only redistributes a Core ML conversion.