carl1758/meetnotes-models

whisper large-v3-turbo Core ML encoder, 8-bit (for whisper.cpp)

A Core ML version of the encoder of OpenAI's whisper large-v3-turbo, with weights palettized to 8 bits. It's for use with whisper.cpp (and whisper.rn) built with Core ML support, next to a ggml-large-v3-turbo*.bin model, so the encoder runs on the Apple Neural Engine.

Compared with the fp16 encoder published in ggerganov/whisper.cpp (ggml-large-v3-turbo-encoder.mlmodelc.zip), the weights are half the size (637,609,152 vs 1,273,969,152 bytes). On a Mac, the Neural Engine memory it holds drops from about 1,247 MB to 646 MB, which lets it run on 4 GB iPhones.

File

ggml-large-v3-turbo-encoder.mlmodelc.zip: unzips to the ggml-large-v3-turbo-encoder.mlmodelc/ folder (the name whisper.cpp looks for next to ggml-large-v3-turbo-q5_0.bin).

• Size: 570,709,613 bytes
• SHA-256: 5b06af945696ec364544e1c9633224f6c381a2ca89bebc61097011e8b16369c7
• MD5: 89303dd190c60078a8716767c2b8e943

How it was built

• Converter: whisper.cpp v1.9.3's own models/convert-whisper-to-coreml.py (--encoder-only True --optimize-ane True --quantize True), with a minimum deployment target of iOS 16 and float32 inputs and outputs.
• Compression: coremltools 9.0 OpPalettizerConfig(mode="kmeans", nbits=8, granularity="per_tensor") on all weight tensors, then compiled with xcrun coremlc compile.
• Pins: torch 2.7.0, openai-whisper 20250625, ane_transformers 0.1.3, Python 3.12.
• Reproducibility: the regenerated fp16 weights are byte-identical to the published fp16 encoder.

Accuracy (Google FLEURS da_dk test, all 930 utterances, whisper.cpp v1.9.3, Apple M4 Pro)

Encoder	WER	CER
ggml (Metal)	14.11 %	4.86 %
Core ML fp16	14.02 %	4.83 %
Core ML 8-bit (this)	13.93 %	4.81 %

The differences are within noise. Built for the meetnotes app (on-device meeting transcription).

License

MIT, following the original whisper weights (© OpenAI).