Research · Speech
Speech to text for Nynorsk: where it stands.
In DeepTranslate tests of five open speech setups from the National Library of Norway, NB-Whisper Large Verbatim transcribed spoken Norwegian into Nynorsk best: 17.8% word error rate on conversation and 14.1% on read speech. Read speech flatters most models.
Published 14 October 2026 · DeepTranslate research
Which model transcribes Nynorsk best?
17.8%
word error rate on conversation, best setup: NB-Whisper Large Verbatim
14.1%
word error rate on read speech, same setup
37.1%
conversation error rate of the standard model, which looks strong on read speech
Why this is hard
Nynorsk is a written standard, not a spoken dialect. People speak Norwegian dialects, and the transcript has to follow Nynorsk spelling.
Most speech models are trained mostly on Bokmål text, so they tend to drift towards it.
Results
Word error rate. Lower is better.
| System | Conversation | Read speech |
|---|---|---|
| NB-Whisper Large Verbatim | Conversation: 17.8% | Read speech: 14.1% |
| NB-Whisper Large | Conversation: 37.1% | Read speech: 12.3% |
| NB-Whisper Large, via Bokmål and automatic conversion | Conversation: 33.0% | Read speech: 16.4% |
| NB-wav2vec2 1B Nynorsk, with language model | Conversation: 29.7% | Read speech: 28.0% |
| NB-wav2vec2 1B Nynorsk, without language model | Conversation: 33.5% | Read speech: 32.7% |
All models are open-source models from the National Library of Norway, each run at a fixed version. Version IDs available on request.
What it means for broadcasters
Read speech flatters a model. The standard NB-Whisper model has the best score on read sentences, but more than triples its error rate on real conversation. Broadcast audio sounds a lot more like conversation.
Writing Bokmål first and converting it to Nynorsk does not fix the problem. The verbatim model, which writes Nynorsk directly, is the most stable across both kinds of speech, and is the one we recommend.
Method and limits
- Two test sets, about one hour in total. Common Voice: volunteers reading sentences aloud, 30 minutes, 32 speakers. NB Samtale (National Library of Norway): unscripted conversation from podcasts and live events, 33 minutes, 19 speakers.
- These are early research results, not a certified benchmark.
- Word error rate counts some correct Nynorsk alternatives as errors, for example Noreg where the reference text says Norge.
- We did not measure how often Bokmål forms replace Nynorsk ones. In our checks this was rare for the recommended model.
- The test sets are not newsroom recordings. Results on broadcast audio may differ.
Frequently asked questions
What is the best speech-to-text model for Nynorsk?
In our tests, NB-Whisper Large Verbatim from the National Library of Norway did best overall: 17.8% word error rate on conversation and 14.1% on read speech.
Does transcribing to Bokmål and converting to Nynorsk help?
No. In our test that route scored 33.0% on conversation, worse than writing Nynorsk directly.