Selbyggen/kb-whisper-small-coreml

KB-Whisper small – WhisperKit / Core ML

KBLab/kb-whisper-small from the National Library of Sweden (Kungliga biblioteket), 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 KBLab's. All credit for the model belongs to KBLab.

Used by the Mac, iPhone and iPad app Sagt for local Swedish transcription («Svensk snabb»).

Conversion

• Tool: whisperkittools (whisperkit-generate-model --model-version KBLab/kb-whisper-small)
• Revision: the default (main, KBLab's «Stage 2»)
• Precision: float16, no quantization
• The conversion's built-in tests (audio encoder and text decoder against the PyTorch reference) passed.
• config.json and generation_config.json are copied unchanged from the original model. The text decoder outputs alignment_heads_weights, so word timestamps work.

Usage with WhisperKit

let folder = try await WhisperKit.download(
    variant: "kb-whisper-small",
    from: "Selbyggen/kb-whisper-small-coreml"
)
let kit = try await WhisperKit(WhisperKitConfig(modelFolder: folder.path))
let result = try await kit.transcribe(audioPath: "inspelning.m4a",
                                      decodeOptions: DecodingOptions(language: "sv"))

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

License

Apache 2.0, the same as the original model.