Speech recognition and models
NB-Whisper
The National Library of Norway's further-trained version of Whisper, adapted to Norwegian speech and dialect.
NB-Whisper is a family of speech recognition models from the National Library of Norway, built by further training Whisper on large amounts of Norwegian speech.
The starting point is the problem that international models have heard little Norwegian. Norwegian is a small language, and a model trained on everything it comes across gets many thousand times more English. The result is that Norwegian works, but as a by-product.
What the further training does
The model gets to hear a lot of Norwegian speech with reference text, from a collection that covers the breadth of dialects. After that it recognises patterns it previously had to guess at: the Bergen r, northern Norwegian intonation, eastern Norwegian contractions.
The effect is largest exactly where Norwegian is hardest. On clear standard eastern Norwegian in good audio, the difference from a general model is small. On broad dialect it is not subtle.
What it does not solve
- Proper nouns. Norwegian surnames and place names remain the largest source of error.
- Nynorsk. Pronunciation does not determine the written standard, and the models generally write bokmål regardless of what is said.
- Code-switching. Norwegian with English jargon mixed in is still demanding. See code-switching.
- Several speakers in the same room. That is a diarisation problem, not a model problem.
Openly available
The models are published openly and can be run locally, in the same way as Whisper. That is why several Norwegian tools build on them.