KB Whisper is a Whisper-family automatic speech recognition model tailored to Swedish. KBLab at the National Library of Sweden developed it for speech-to-text workloads, and its training used more than 50,000 hours of Swedish audio across sources including subtitles, parliamentary speeches, and specialist speech corpora. Training proceeded in two stages: continued pretraining on a broad audio collection, followed by fine-tuning on a smaller, higher-quality set.

The model is best suited to Swedish transcription, especially when Swedish accuracy matters more than general multilingual coverage. A secondary model overview reports a 47% reduction in word error rate compared with OpenAI Whisper across several evaluation datasets, although the underlying benchmark tables are not included in the supplied evidence. Its practical flexibility is supported by available Hugging Face, whisper.cpp GGML, ONNX, and CTranslate2 formats, plus compatibility with faster-whisper and WhisperX workflows. Speaker diarization is not an inherent capability; a multi-step pipeline can combine transcription and alignment when speaker separation is required.