Dutch ASR Leaderboard
An independent, community-driven benchmark for Dutch automatic speech recognition.
Models are evaluated on standardized public test sets; models we run ourselves also on private datasets that are not released, so they can't have been trained on. Lower WER is better. Rankings serve as a proxy for comparison, performance on your data may differ.
Note: Some models may be benchmaxxed: trained or fine-tuned on data that overlaps with these test sets. Treat results as indicative, not definitive. Submit your model on GitHub →
Public WER / CER: average over FLEURS, VoxPopuli and MLS, open test sets that every model is run on. Private WER / CER 🔒: average over private datasets that are not released, so no model can have been trained on them. Each covers a domain; for now one: Medical. Only models we run on our own hardware have a private score: an API would receive the audio, so API models show —. The two are never averaged together.
Leaderboard chooses which average ranks the table (⬇️) and shows its WER per dataset (public) or per domain (private); models without that score are listed unranked at the bottom. Sort by another column with the ⋮ menu next to its name. WER and CER in %, lower is better; best per column highlighted. Speed: processing time divided by audio duration, averaged over the public datasets; lower is faster. For API models it includes network time, so it is not comparable with local models.