Improving Phoneme Recognition with Augmented Autoregressive Predictive Coding

Self-supervised learning (SSL) has improved the performance in various speech processing downstream tasks. Pretraining speech representations with SSL models requires large datasets which can be hard to collect for low resource language scenarios. To address this problem we propose to use audio augmentation to pre-train the SSL model with a smaller dataset but augmented. We illustrate our technique using autoregressive predictive coding (APC) by using augmentation effects directly at an audio level in order to generate synthetic training data. As a result, we aim to increase robustness in trained SSL model. We apply two kinds of audio augmentation, speed perturbation and speed perturbation with pitch change. We then evaluate our pre-trained APC model on a downstream phoneme recognition task. From our results and analysis, we conclude that applying augmentation in a low resource configurations can improve the robustness in the pre-trained model resulting in improved phoneme classification accuracy for a downstream task.