The Use of Machine Learning Algorithms in the Classification of Sound: A Systematic Review

Introduction

Sound or acoustic signals are gradually gaining research popularity as a tool for environmental monitoring, security surveillance, diagnoses of diseases, critical information infrastructure protection, and data transmission (Bourouhou et al., 2019; Ibrahim et al., 2018; Loey et al., 2020; Luque et al., 2018). Sound is considered as the second most important sense after sight that is capable of carrying information about the environment (Perr, 2005). Although sound varies depending on seasons, time, geographic location as well as propagation medium, it is considered as one of the most significant signals used to monitor and detect changes in the environment. Accordingly, the ability to differentiate (classify) one sound or acoustic signal type from another is pertinent that, if accomplished, would result in significant progress in application areas such as early warning disaster management, medical diagnosis (Loey, Naman, & Zayed, 2020), and action or event detection. Recent studies have shown that machine learning (ML) algorithms are efficient in the domains of image and speech recognition, natural language processing, medical imaging, data extraction (Dwivedi et al., 2019; Malfante et al., 2018; Tatoian & Hamel, 2018) and text classification (Elfergany & Adl, 2020; Sangwan & Bhatnagar, 2020).

Classification aims at predicting accurately the target object and differentiating one object class from the other given a set of data. It is predominantly performed using selected features that feed classifier tools such as machine learning and neural networks (Mitilineos et al., 2018). In particular, sound classification is aimed at classifying audio segments into specific classes which requires the understanding of the fundamental structure of frequencies in acoustic signals (Dwivedi et al., 2019). This is commonly addressed with features used in speech and music processing such as MFCC (Mel frequency cepstral coefficient), linear prediction coefficients (LPC), linear prediction cepstral coefficients (LPCC) and fast Fourier transforms (Briggs et al., 2012; Chu et al., 2009; Davis & Suresh, 2019; Karbasi et al., 2011; Mitilineos et al., 2018; Oletic et al., 2012; Pramono et al., 2017; Sengupta et al., 2016). A variety of machine learning techniques have also been adopted to obtain robust sound classification models.

Considering the plethora of acoustic features and machine learning (ML) algorithms coupled with the nature of sound, it is imperative to offer researchers an indication of the major research trends and methodologies that can assist in designing and developing automatic sound classification systems. Accordingly, this study provides summaries of the existing literature on algorithms for the classification of sound and analyzes the use of ML in the various sound classification tasks. The specific objective of the review is to identify: (a) publication patterns in acoustic signal classification, (b) trends in the use of ML in acoustic signal/sound classification, (c) open questions and challenges in the use of ML algorithms in acoustic signal classification, and (d) research gaps in the subject area.