How accurate is Norwegian speech to text with local Whisper?
In the Whisper paper itself (arXiv:2212.04356), the large-v2 model scores 9.5% word error rate on Norwegian FLEURS. Independent Norwegian speech research corroborates that order of magnitude, measuring OpenAI's large-v3 at 10.4% on FLEURS Bokmål and 6.8% on the NST Bokmål corpus. So roughly 7 to 10% on clean read Bokmål is the honest range.
Every one of those numbers describes Bokmål, and that is the crucial detail rather than a footnote. See the next section.
Accuracy tracks model size far more than it tracks local versus cloud - it is the same open-source Whisper either way. NB-Whisper, the Norwegian fine-tune from the National Library of Norway, scores better than base Whisper; we ship the base model, so the figures above are what to expect.
Two written standards, and Whisper effectively knows one
Norwegian has two official written standards, Bokmål and Nynorsk. Around 13% of Norwegian schoolchildren write Nynorsk. Base Whisper effectively produces Bokmål, and on Nynorsk it degrades sharply: Norwegian ASR research measured OpenAI Whisper large at 30.0% word error rate on Common Voice Nynorsk, and found that smaller model sizes exceed 100% word error rate on it. For comparison, a Norwegian fine-tune of the same architecture reached 12.6% on that benchmark - but that is their model, not ours, and we are not going to borrow its numbers.
We expose no control over which standard comes out. There is no Bokmål-versus-Nynorsk setting in SnailText, because Whisper offers no such selection. If your work requires consistent Nynorsk - and for some Norwegian public-sector and municipal work it genuinely does - this tool will not deliver it.
There is no spoken standard, and dialects are used everywhere. Norwegian dialects differ substantially in grammar, pronunciation and vocabulary, and are used in formal contexts including parliament and broadcast news. One study of Norwegian dialects recorded 38 distinct pronunciations of the single question word meaning who. A further wrinkle: the written norms themselves permit substantial variation, so two different transcriptions can both be correct, which means measured word error rates understate perceived quality and the model's output may simply not match your personal spelling habits.
What we actually offer for Norwegian, and what we do not
For Spanish, German, French, Portuguese and Dutch, SnailText ships a language-specific prompt. For Norwegian we do not. There is no Norwegian tuning layer, no Norwegian cleanup pass and no Norwegian fine-tune, and no way to request a particular written standard.
What SnailText does give you is the best open Norwegian speech model running entirely on your own hardware, with a global hotkey that pastes text at your cursor in any app, and audio that never leaves your machine. For Bokmål dictation that is a solid combination.
If you write Nynorsk, the honest recommendation is to look at NB-Whisper, the Norwegian fine-tune published by the National Library of Norway, rather than this. We would rather point you there than take the download and have you find out afterwards.
Set the dictation language to Norwegian explicitly rather than leaving it on automatic. Whisper decides the language from roughly the first thirty seconds, and written Norwegian and Danish can be extremely close.