whisper-large-v3-ivrit-coreml

whisper-large-v3-ivrit-coreml

CoreML encoder (.mlmodelc) conversion of ivrit-ai/whisper-large-v3 for use with whisper.cpp on Apple Silicon. The encoder runs on the Apple Neural Engine (ANE) when present, freeing Metal/CPU for the decoder.

Usage with whisper.cpp

1. Download ivrit-ai-whisper-large-v3-encoder.mlmodelc.zip and the original GGML weights (.bin) from ivrit-ai/whisper-large-v3-ggml.
2. Place them side-by-side, with the .mlmodelc directory named to match the .bin filename (replacing .bin with -encoder.mlmodelc):

~/Your-Models-Dir/
├── ivrit-ai-whisper-large-v3.bin
└── ivrit-ai-whisper-large-v3-encoder.mlmodelc/   # ← extracted from this repo's zip

1. whisper.cpp v1.8.4+ (built with WHISPER_USE_COREML=1, which is the default in the official xcframework) auto-detects the sibling .mlmodelc on whisper_init_from_file_with_params and routes encoder forward passes to CoreML.

On first load, CoreML compiles the model for the current device (~10–15 s); every subsequent load is instant.

Conversion details

• Generated via whisper.cpp's models/convert-h5-to-coreml.py --model-name large-v3 --encoder-only True --optimize-ane True against the ivrit-ai HuggingFace checkpoint.
• 100% of encoder weights came from the ivrit-ai fine-tune (no fallback to OpenAI's base large-v3).
• Validated against the HuggingFace reference encoder on random log-mel inputs: median per-position diff ≈ 0.02, p99 diff ≈ 0.31 — within FP16 quantization noise.
• Input shape: (1, 128, 3000) (whisper large-v3 standard).
• Output shape: (1, 1500, 1280).
• Compute precision: Mixed FP16/FP32/Int32, ANE-optimized.

License

Apache 2.0 — same as the upstream ivrit-ai/whisper-large-v3 base model. The base model is itself a fine-tune of openai/whisper-large-v3 (MIT).

This conversion is a derivative work. Please retain the attribution chain when redistributing.

Acknowledgements

• ivrit-ai for the Hebrew fine-tuning of Whisper large-v3.
• OpenAI for the original Whisper architecture and weights.
• ggerganov / whisper.cpp for the conversion tooling.