nb-whisper-small — WhisperKit CoreML

NbAiLab/nb-whisper-small converted to CoreML for WhisperKit, so it runs on the Apple Neural Engine.

This is a format conversion, not a new model. No retraining, no fine-tuning, no distillation. Weights and behaviour are NbAiLab's; the changes are the PyTorch → CoreML conversion and the folder layout WhisperKit expects. All credit for the model belongs to NbAiLab.

Contents

nb-whisper-small/
├── MelSpectrogram.mlmodelc     372 KB
├── AudioEncoder.mlmodelc       170 MB
├── TextDecoder.mlmodelc        293 MB
├── config.json
└── generation_config.json

Float16 activations. The layout mirrors argmaxinc/whisperkit-coreml, which is what WhisperKitConfig(model:modelRepo:) expects to find.

Usage

With papegøye, a bilingual Norwegian/English macOS dictation daemon — it downloads this repo automatically:

parrot --bilingual

Directly with WhisperKit:

let config = WhisperKitConfig(
    model: "nb-whisper-small",
    modelRepo: "Barrymanalow/nb-whisper-coreml"
)
let pipe = try await WhisperKit(config)

Requires macOS 14+ on Apple Silicon. Verified with WhisperKit 0.9 on macOS 15; loads and transcribes Norwegian correctly.

Conversion

Converted with whisperkittools via scripts/convert-nb-whisper.sh in the papegøye repo. Two things that trip up a naive run, both handled by that script:

• whisperkittools pins torch==2.5.0, which has no wheels above CPython 3.12. On a Mac defaulting to a newer Python, pip install fails with a bare "no matching distribution" that never mentions the version.
• whisperkit-generate-model emits only the compiled .mlmodelc bundles. config.json and generation_config.json stay behind in the source snapshot and have to be copied across, or WhisperKit has no config to read.

Reproduce with:

scripts/convert-nb-whisper.sh small

Caveats

It is a multilingual model, not a Norwegian-only one. NB-Whisper is Whisper small fine-tuned on Norwegian, so the multilingual tokenizer is intact (vocab_size 51865). Practical consequences:

• English words inside Norwegian speech are transcribed rather than mangled.
• Forcing the <|no|> token onto English audio does not fail loudly — it renders the English speech as fluent Norwegian, which reads like a correct translation. Anything routing audio to this model should be confident the audio really is Norwegian. See papegøye#22.

Not verified here: Apple Neural Engine residency under powermetrics, and a systematic word-level diff against the PyTorch checkpoint. The CoreML compute plan targets ANE for nearly every operation, but that is the plan, not a measurement.

License

Apache-2.0, inherited from NbAiLab/nb-whisper-small. The conversion adds no restrictions. If you use this, cite NbAiLab's work — not this repo, which only changed the file format.