Effective Masking Shapes Based Robust Data Augmentation for Acoustic Scene Classification

Abstract

In this paper, we explore diverse masking shapes for en-hancements over the traditional SpecAugment technique, which predominantly employs a cross-shaped masking, for improved performance in acoustic scene classification (ASC). The importance of data augmentation, particularly masking, is well acknowledged in the field of ASC for increasing the robustness and accuracy of the deep learning models. We examine the effects of various masking shapes such as rect-angles and circles, and the rate at which they were applied during batch processing using the ASC dataset. A meticulous series of experiments was conducted in which different masking shapes were implemented with varying application rates in each batch. This leads to extensive evaluations of classification performance. These results can be instrumental in refining data augmentation strategies through masking, contributing to the development of more potent and efficient systems for ASC.