kb-whisper-large

Overview. KB-Whisper Large is a Swedish automatic-speech-recognition model built by KBLab, the research unit of the National Library of Sweden. It is a 1.55B-parameter encoder-decoder fine-tune of Whisper large-v3 that produces text from audio, and it was trained through two-stage continued pretraining plus finetuning on roughly 56,500 hours and then 8,533 hours of Swedish speech. A developer reaches for it when the job is Swedish transcription and the quality of a general multilingual ASR model is not good enough. It is released under Apache-2.0 with publicly downloadable weights.

Key strengths. The model's reported word error rates are 5.4 on FLEURS, 4.1 on CommonVoice, and 5.2 on NST, against 7.8, 9.5, and 11.3 for the Whisper large-v3 baseline it was built from, a substantial reduction across all three Swedish evaluation sets. Its two-stage training on a large body of Swedish speech is the reason for that gap, since the base model was not specialised for the language. KBLab also documents three Stage-2 transcription styles (standard, the default; subtitle for more condensed output; strict for more verbatim output), so output style can be matched to the task.

When to use it. Use it for Swedish audio where transcription accuracy matters more than multilingual coverage: recorded meetings, interviews, broadcast material, and public-sector archives in Swedish. The strict style suits cases that need a verbatim record, while subtitle is the better fit when transcripts feed downstream display or captioning. Because the model is open weights under Apache-2.0, it is also a candidate when the deployment itself must run on infrastructure you control rather than a vendor's.

Limitations. The model is Swedish only, so it is not a general multilingual ASR choice. It has a text output cap of 448 tokens per 30-second audio window, and it has no conventional token context window, no vision input, no audio output, no tool or function calling, and no reasoning-effort controls. No training-data cutoff date is published.

Sovereignty note. KB-Whisper Large is built by KBLab at the National Library of Sweden and released as open weights under Apache-2.0, so it can be self-hosted. Data residency and retention for hosted calls are properties of the serving endpoint and are shown per endpoint on this page.