We investigate the possibility of forcing a self-supervised model trained using a contrastive predictive loss, to extract slowly varying latent representations. Rather than producing individual predictions for each of the future representations, the model emits a sequence of predictions shorter than the sequence of upcoming representations to which they will be aligned. In this way, the prediction network solves a simpler task of predicting the next symbols, but not their exact timing, while the encoding network is trained to produce piece-wise constant latent codes. We evaluate the model on a speech coding task and demonstrate that the proposed Aligned Contrastive Predictive Coding (ACPC) leads to higher linear phone prediction accuracy and lower ABX error rates, while being slightly faster to train due to the reduced number of prediction heads.
Cite as: Chorowski, J., Ciesielski, G., Dzikowski, J., Łańcucki, A., Marxer, R., Opala, M., Pusz, P., Rychlikowski, P., Stypułkowski, M. (2021) Aligned Contrastive Predictive Coding. Proc. Interspeech 2021, 976-980, doi: 10.21437/Interspeech.2021-1544
@inproceedings{chorowski21b_interspeech,
author={Jan Chorowski and Grzegorz Ciesielski and Jarosław Dzikowski and Adrian Łańcucki and Ricard Marxer and Mateusz Opala and Piotr Pusz and Paweł Rychlikowski and Michał Stypułkowski},
title={{Aligned Contrastive Predictive Coding}},
year=2021,
booktitle={Proc. Interspeech 2021},
pages={976--980},
doi={10.21437/Interspeech.2021-1544}
}