Modelpublic

sebastian-duadigital/whisperkit-kb-whisper-small

sourceHugging Faceapache-2.0updated 11d agoView on Hugging Face

0likes77downloads

Model Card

KB-Whisper Small for WhisperKit (CoreML)

A CoreML conversion of KBLab/kb-whisper-small, the Swedish-tuned Whisper small model from the National Library of Sweden, packaged in the folder layout that WhisperKit expects. It exists so that Dikta, an offline dictation app for macOS, can download a Swedish model it controls.

What is in this repository

KBLab_kb-whisper-small/
  AudioEncoder.mlmodelc/
  MelSpectrogram.mlmodelc/
  TextDecoder.mlmodelc/
  config.json
  generation_config.json

About 464 MB in total. config.json and generation_config.json are copied verbatim from the source checkpoint. No tokenizer files are included; WhisperKit resolves the Whisper tokenizer from its default repository.

Provenance

Item	Value
Source checkpoint	`KBLab/kb-whisper-small`, revision `3564d61a42fc210ceaa55a22a96dd64478959c78` (Stage 2 default)
Converter	Argmax `whisperkittools`, commit `84f77a83c8f530022ae55fbb1a64b3351ef63c7a`
Conversion host	Apple M2 Max, macOS 26.6, Xcode 27.0, Python 3.11, torch 2.5.0, coremltools 9.0
Reproduction	`dikta-macos/bench/convert/convert.sh` in the Dikta repository, which downloads the pinned revision and runs the converter end to end

Weights are unchanged from the source; only the file format differs.

Measured accuracy

Word error rate on a fixed, seeded 20-clip sample per language from the FLEURS test split, scored with jiwer after lowercasing and punctuation stripping. Same harness for every row.

Model	Swedish	English
This model	3.5%	54.2%
`openai_whisper-large-v3-v20240930_turbo_632MB`	10.1%	6.7%
`openai_whisper-small`	18.5%	9.9%

Use it for Swedish only. It is not a general multilingual model; English output is poor.

Usage with WhisperKit

swift

let pipe = try await WhisperKit(
    model: "KBLab_kb-whisper-small",
    modelRepo: "sebastian-duadigital/whisperkit-kb-whisper-small"
)
let result = try await pipe.transcribe(audioPath: "clip.wav", decodeOptions: .init(language: "sv"))

License and credit

The weights are released by KBLab under Apache 2.0, and this conversion keeps that license. If you use the model in research, cite KBLab's paper:

Vesterbacka, L., Rekathati, F., Kurtz, R., Sikora, J., Toftgård, A. (2025)
Swedish Whispers; Leveraging a Massive Speech Corpus for Swedish Speech Recognition.
Proc. Interspeech 2025, 758-762, doi: 10.21437/Interspeech.2025-2646