Selbyggen/nb-whisper-medium-coreml

NB-Whisper medium – WhisperKit / Core ML

NbAiLab/nb-whisper-medium from the National Library of Norway (Nasjonalbiblioteket), converted to Core ML for WhisperKit so it runs on the Apple Neural Engine.

This is a format conversion, not a new model. The weights and behaviour are the National Library's. All credit for the model belongs to NbAiLab.

Used by the Mac, iPhone and iPad app Sagt for local Norwegian transcription («Norsk nøyaktig»).

Conversion

• Tool: whisperkittools (whisperkit-generate-model --model-version NbAiLab/nb-whisper-medium)
• Revision: main (commit 0ed074d5985bd56ca4140159a9dbffbc3fb5117e)
• Precision: float16, no quantization (folder NbAiLab_nb-whisper-medium)
• config.json and generation_config.json are copied unchanged from the original model.

Accuracy

Word error rate on real Norwegian speech, measured with the same clips, normalisation and code path on the Neural Engine. In our measurements NB-Whisper medium was as accurate as NB-Whisper large on every set, at half the size:

Set	NB-Whisper medium (this)	NB-Whisper large
FLEURS nb (75 clips)	6.5 %	7.0 %
NST (75 clips)	2.6 %	2.2 %
NB Tale, spontaneous speech (75 clips)	15.8 %	17.1 %
NPSC, Storting (75 clips)	14.6 %	14.8 %
Storting, four ~10-minute meetings	13.6 %	13.9 %

The original PyTorch model gives the same numbers within noise. NPSC and NB Tale have verbatim references, while NB-Whisper writes clean text (drops repetitions and false starts), which explains most of the error there.

Usage with WhisperKit

let folder = try await WhisperKit.download(
    variant: "NbAiLab_nb-whisper-medium",
    from: "Selbyggen/nb-whisper-medium-coreml"
)
let kit = try await WhisperKit(WhisperKitConfig(modelFolder: folder.path))
let result = try await kit.transcribe(audioPath: "opptak.m4a",
                                      decodeOptions: DecodingOptions(language: "no"))

The model uses the same tokenizer as openai/whisper-medium.

License

Apache 2.0, the same as the original model.