---
title: "Whisper Scandinavian Dialect ASR Benchmark"
url: https://ypai.ai/research/asr-benchmark/
description: "Whisper large-v3, medium and small word error rate across six Scandinavian dialects, Standard Bokmål to Jutlandic Danish. Methodology, raw data, CSV download."
source: "src/copy/routes (route /research/asr-benchmark/)"
---

# Whisper Scandinavian Dialect ASR Benchmark

> Whisper large-v3, medium and small word error rate across six Scandinavian dialects, Standard Bokmål to Jutlandic Danish. Methodology, raw data, CSV download.

## Whisper Norwegian and Scandinavian Dialect ASR Benchmark

YPAI Research (2026). Whisper Norwegian and Scandinavian Dialect ASR Benchmark. YPAI. https://ypai.ai/research/asr-benchmark/ @misc{ypai2026whisperscandinavian, author = {YPAI Research}, title = {Whisper Norwegian and Scandinavian Dialect ASR Benchmark}, year = {2026}, url = {https://ypai.ai/research/asr-benchmark/}, note = {CC-BY 4.0}, }

OpenAI Whisper achieves 8 to 11% Word Error Rate on Standard Bokmål. On Northern Norwegian dialect, the same model exceeds 40% WER. Below is the full per-dialect benchmark across six Scandinavian varieties, three Whisper model sizes, with raw data downloadable as CSV.

## Key findings

- Whisper large-v3 WER climbs from 8-11% on Standard Bokmål to 34-42% on Northern Norwegian dialect, a 3-4x degradation on a model marketed as multilingual.
- Language ID misfires on dialectal speech: Western Norwegian is misidentified as Swedish 31% of the time, doubling effective WER when language forcing is not applied.
- Under automotive in-cabin conditions (55-72 dB road noise, 2-5 word utterances), no Whisper model size achieves usable WER outside Standard Bokmål.
- 50-200 hours of dialect-native conversational speech, with native-speaker annotation, reduces WER by 30-45% relative. Read speech alone, regardless of volume, does not close the gap.

## Results: Word Error Rate by Dialect

All values are 95% confidence intervals across multiple test runs. WER measured against normalized Bokmål reference transcriptions for Norwegian; standard orthography for Swedish and Danish.

Read the full methodology & analysis

## Methodology

Six Scandinavian dialect groups evaluated on a minimum of 4 hours of audio per group. Corpus size per dialect ranged from 4.2 to 7.8 hours. Speaker counts ranged from 18 to 34. Audio captured in controlled studio conditions and supplemented with in-cabin automotive test recordings under road noise at 55-72 dB SPL.

Whisper large-v3, medium, and small models tested with language forcing applied (

). WER calculated against normalized Bokmål reference transcriptions for Norwegian dialect groups, with a secondary phonetic tier for dialect-specific forms.

Annotator agreement averaged 91.3% for Standard Bokmål, dropping to 78.6% for Northern Norwegian and 74.1% for Jutlandic Danish. Disagreements adjudicated by a third dialect-specialist annotator.

Test corpus is not publicly distributable due to GDPR Article 9 biometric data restrictions on Nordic speech recordings. Aggregate WER results, methodology, and dialect-specific phoneme analysis are released under CC-BY 4.0 for academic and commercial reference.

## How to cite

## Need ASR that works on your target dialects?

YPAI builds dialect-native speech corpora for enterprise ASR: native annotators, deployment-realistic acoustic conditions, GDPR + EU AI Act compliant consent frameworks. Talk to us about your language coverage requirements.