Aligned Contrastive Predictive Coding
This work addresses the challenge of improving representation learning for speech processing, though it appears incremental as it builds on existing contrastive predictive methods with a specific alignment modification.
The paper tackled the problem of extracting slowly varying latent representations in self-supervised learning by proposing Aligned Contrastive Predictive Coding (ACPC), which simplifies prediction tasks and results in higher linear phone prediction accuracy and lower ABX error rates on a speech coding task.
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 that of the 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.