Data Augmentation and Deep Learning Methods in Sound Classification: A Systematic Review

Abstract

The aim of this systematic literature review (SLR) is to identify and critically evaluate current research advancements with respect to small data and the use of data augmentation methods to increase the amount of data available for deep learning classifiers for sound (including voice, speech, and related audio signals) classification. Methodology: This SLR was carried out based on the standard SLR guidelines based on PRISMA, and three bibliographic databases were examined, namely, Web of Science, SCOPUS, and IEEE Xplore. Findings. The initial search findings using the variety of keyword combinations in the last five years (2017–2021) resulted in a total of 131 papers. To select relevant articles that are within the scope of this study, we adopted some screening exclusion criteria and snowballing (forward and backward snowballing) which resulted in 56 selected articles. Originality: Shortcomings of previous research studies include the lack of sufficient data, weakly labelled data, unbalanced datasets, noisy datasets, poor representations of sound features, and the lack of effective augmentation approach affecting the overall performance of classifiers, which we discuss in this article. Following the analysis of identified articles, we overview the sound datasets, feature extraction methods, data augmentation techniques, and its applications in different areas in the sound classification research problem. Finally, we conclude with the summary of SLR, answers to research questions, and recommendations for the sound classification task.

Detecting Lombard Speech Using Deep Learning Approach

“…In machine learning-based processing, this is carried out by incrementing training data. A standard solution is to artificially increase the quantity of training data patterns by transforming the available speech patterns by adding noise, time warping and shifting, pitch shifting, time or frequency masking, or filtering [ 58 , 59 , 60 ].…”

Sound-Event Detection of Water-Usage Activities Using Transfer Learning

“…Once the clip length was fixed, we set the frame duration to 1 s, considering the standard frame size in YAMNet input, and ensured that adjacent frames had a 50% overlap. Through experiments, we concluded that 3 s is an appropriate duration [ 22 ].…”

Robotics Perception and Control: Key Technologies and Applications

“…It outperforms traditional classification methods in handling real-world industrial mechanical sound data, thereby contributing to reduced maintenance costs, enhanced safety in processing, improved equipment availability, and reduced production downtime costs while maintaining acceptable performance levels. However, this deep learning method requires extensive data when dealing with complex audio signals and industrial noise, or its performance may be compromised [44].…”