A Dutch Wav2Vec2-base model, pre-trained on 831 hours of exclusively Dutch speech.
More information, incl. the training manifest and configuration is available in the
Wav2Vec2-NL repository on Zenodo.
Analyses of Dutch phonetic and lexical features encoded in Wav2Vec2-NL hidden states are reported in the paper
What do self-supervised speech models know about Dutch? Analyzing advantages of language-specific pre-training (Interspeech 2025; see full citation
below).
Note: This model does not have a tokenizer as it was pretrained on audio alone. In order to use this model for speech recognition, a tokenizer should be created and the model should be fine-tuned on labeled text data. Check out
this blog for an explanation of fine-tuning Wav2Vec2 models on HuggingFace.
1from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2Model
2
3feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained('amsterdamNLP/Wav2Vec2-NL')
4model = Wav2Vec2Model.from_pretrained('amsterdamNLP/Wav2Vec2-NL')
The
Wav2Vec2-NL model was published as part of:
de Heer Kloots, M., Mohebbi, H., Pouw, C., Shen, G., Zuidema, W., Bentum, M. (2025). What do self-supervised speech models know about Dutch? Analyzing advantages of language-specific pre-training.
Proc. INTERSPEECH 2025.
https://doi.org/10.48550/arXiv.2506.009811@inproceedings{deheerkloots25_interspeech,
2 title = {What do self-supervised speech models know about Dutch? Analyzing advantages of language-specific pre-training},
3 author = {Marianne {de Heer Kloots} and Hosein Mohebbi and Charlotte Pouw and Gaofei Shen and Willem Zuidema and Martijn Bentum},
4 year = {2025},
5 booktitle = {Interspeech 2025},
6 doi = {10.21437/Interspeech.2025-1526},
7}
