Audio Data Augmentation for Environmental Sound Classification

Abstract

Audio Data Augmentation for Environmental Sound Classification Abstract: Convolutional neural networks (CNNs) are effective for extracting features and sound classification, however real-time sound is affected by past sequences. Furthermore, the primary limitation of deep learning techniques is their requirement for an enormous quantity of datasets in order to demonstrate their effective operation. This research focusses on using an RNN (recurrent neural network) in conjunction with CNN to tackle this issue. Furthermore, for high-quality data augmentation, a DCGAN (Deep Convolutional Generative Adversarial Network) is employed. To expand the labeled training dataset, data augmentation is used. The technique of data augmentation enhances sound classification efficiency. Three types of augmentation methods are used like Traditional data augmentation, Traditional data augmentation with specifications and DCGAN based approach. The effectiveness of these data augmentation techniques on the Urbansound8K audio dataset are examined. The amount of data samples are increased into numerous by using these augmentation methods. According to the results, DCGAN approach can produce spectrograms and increase classification accuracy better as compared to traditional approaches. The images it creates possess characteristics that are comparable to those of the original training images.