KB-Whisper large, MLX fp16
An unofficial MLX conversion of KBLab/kb-whisper-large by KBLab (the National Library of Sweden's data lab), for running
locally on Apple Silicon with mlx-whisper. Published by the
Tibber-se organisation as a convenience for colleagues and the community. It is not affiliated with or
endorsed by KBLab, and comes with no support. All credit for the model goes to them.
Use
uvx --from mlx-whisper mlx_whisper recording.m4a --model Tibber-se/kb-whisper-large-mlx --language sv -f srt
Needs an Apple Silicon Mac and ffmpeg. The weights stay on your machine and so does the audio.
Provenance
- Source:
KBLab/kb-whisper-largeat revisiond5d5984b4d8f7c4847a8ea203f1976285fb28300(Apache-2.0). - Converter: 796f5b5 (ml-explore/mlx-examples, whisper/convert.py), run locally,
--dtype float16. - Files:
weights.safetensors(named for themlx_whisperloader on PyPI) andconfig.json. Same tensors as the converter's ownmodel.safetensorsoutput.
How it was checked
Word error rate on 100 clips of the FLEURS test split (Swedish, CC-BY-4.0), one run, on an Apple Silicon Mac:
| WER | |
|---|---|
whisper-large-v3-turbo (MLX) |
8.7% |
KBLab/kb-whisper-large through transformers (fp32, original weights) |
4.9% |
| This model | 5.1% (median clip 2.7%, worst clip 33%) |
Corpus WER is within about 0.3 points of the original weights. FLEURS is clean read speech, so real meetings and calls will score worse. Normalisation lowercases and strips punctuation but does not normalise numbers. A single run on 100 clips per language is a sanity check, not a benchmark.
Limitations
- Word-level timestamps were not tested and probably do not work: the converter only sets alignment heads for the official OpenAI models.
- Whisper models can hallucinate text on silence, noise or music. Check transcripts that matter.
- Large size only. Other sizes were not converted.
- If you process recordings of people, you are responsible for having a legal basis to do so.