Large SSL Models for Haitian Creole

Wav2Vec2 and Data2Vec-Audio self-supervised models of speech processing for Haitian Creole (HAT 1.4K)

Posted: July 02, 2025 1 min.

The HAT 1.4K models are monolingual self-supervised models of speech processing for Haitian Creole (Kreyòl), pre-trained from scratch on approximately 1,400 hours of speech. Three models are available:

These models position Haitian Creole as a high-resource language in terms of self-supervised models of speech processing: fine-tuned on an automatic speech recognition task, they achieve performances comparable to, or even surpassing, larger models derived from large-scale multilingual pretraining (XLSR-53, XLSR2-300M, MMS-1B). They are described in the following papers (Havard et al., 2025) and (Havard et al., 2025), and can be adapted to other French-based Caribbean Creoles.

The code used to pre-train these models is available on GIN: CREAM/SSL-Haitian. The models were pre-trained on several corpora of Haitian Creole, including the Atlas linguistique d’Haïti and the Radio Haiti archives.

Access to the models is granted upon request. If you use these models in your own work, please cite our Interspeech paper.

References

  1. Havard, W. N., Govain, R., Lecouteux, B., & Schang, E. (2025). Self-Supervised Models of Speech Processing for Haitian Creole. Interspeech 2025, 4018–4022. https://doi.org/10.21437/Interspeech.2025-1852

  2. Havard, W. N., Govain, R., Lecouteux, B., & Schang, E. (2025). Modèles auto-supervisés de traitement de la parole pour le Créole Haïtien. Actes de CORIA-TALN-RJCRI-RECITAL 2025. Actes Des 32ème Conférence Sur Le Traitement Automatique Des Langues Naturelles (TALN), Volume 1 : Articles Scientifiques Originaux, 542–554. https://talnarchives.atala.org/TALN/TALN-2025/98.pdf

Share on