nb-whisper-large

Overview. NB-Whisper Large is an open-weights speech-recognition and speech-translation model from NB AI-Lab at the National Library of Norway, built by fine-tuning OpenAI's Whisper architecture and released under Apache-2.0. It converts audio into transcribed or translated text, with a focus on Norwegian and broader multilingual coverage. A developer would reach for it when the task is accurate Norwegian transcription of real recordings rather than text generation.

Key strengths. The peer-reviewed NB-Whisper paper reports a word error rate of 6.6 on FLEURS Bokmål versus 10.4 for Whisper Large-v3, and 2.2 on NST versus 6.8, so the fine-tune measurably improves Norwegian spoken-language accuracy over its base. It is a 1.55B-parameter dense encoder-decoder that keeps Whisper's multilingual reach while specialising on Norwegian corpora such as NCC, NST, NPSC, NRK subtitles, Stortinget speeches and audiobooks. The served endpoint exposes word-level timestamps and speaker diarisation, so transcripts can be aligned to audio and attributed to speakers without a separate pipeline.

When to use it. Use this model for Norwegian audio where transcription accuracy matters more than interactive latency: interviews, parliamentary recordings, broadcasts, and audiobook-length material. The served endpoint transcribes at roughly 15 to 25 times real time, which suits batch jobs over long recordings; word-level alignment adds overhead and is best enabled only when timestamps are actually needed. Speaker diarisation makes it a fit for multi-speaker material such as meetings or debates.

Limitations. This is an ASR model, not a text assistant: it has no tool calling, structured output, JSON mode, or reasoning traces, and it does not perform document understanding. Whisper processes audio in fixed 30-second windows with a 448-token decoder output per window, so very long inputs are handled as sequences of windows rather than a single context. The base-checkpoint provenance is slightly uncertain (the model card's tag and prose disagree between Whisper Large and Large-v3), no training-data cutoff is published, and the licence permits reuse but the exact fine-tune variant is not formally documented.

Sovereignty note. The model is created by NB AI-Lab in Norway and served on Stockholm infrastructure that the vendor states sits outside US CLOUD Act and FISA 702 reach, with open Apache-2.0 weights available for self-hosting. The vendor states it retains no prompt or output content and never trains on customer data, though no audio-endpoint-specific retention period is published, so data-handling behaviour should be confirmed with the endpoint operator before routing sensitive audio.