smkrv/whisper-podlodka-turbo-coreml

Whisper Podlodka Turbo — CoreML (WhisperKit)

CoreML version of bond005/whisper-podlodka-turbo converted for use with WhisperKit on Apple Silicon.

Model Details

• Architecture: Whisper Turbo (openai/whisper-large-v3-turbo fine-tuned)
• Format: CoreML .mlmodelc (compiled, ready for Apple Neural Engine)
• Size: ~1.5 GB total
• Languages: Russian (primary), English
• Fine-tuning: Trained on Russian podcast data (Podlodka podcast + Golos dataset)

Model Files

bond005_whisper-podlodka-turbo/
├── AudioEncoder.mlmodelc/     — Audio encoder (1.2 GB)
├── TextDecoder.mlmodelc/      — Text decoder (328 MB)
├── MelSpectrogram.mlmodelc/   — Mel spectrogram preprocessing
├── config.json                — WhisperKit configuration
├── generation_config.json     — Generation parameters
├── vocab.json, merges.txt     — BPE tokenizer
├── tokenizer_config.json
├── normalizer.json
└── *.mlcomputeplan.json       — CoreML compute plans

Usage

High-quality Russian ASR model for WhisperKit on Apple Silicon.

// WhisperKit integration
let pipe = try await WhisperKit(
    modelFolder: "/path/to/bond005_whisper-podlodka-turbo",
    download: false
)
let result = try await pipe.transcribe(audioPath: audioURL.path)

Attribution

Base model whisper-large-v3-turbo by OpenAI. Russian fine-tune whisper-podlodka-turbo by @bond005. CoreML conversion for WhisperKit by @smkrv.