Sound Identification using MFCC with Machine Learning

This article presents an application of sound identification using machine learning techniques. Identification for access control system such as an entering-exiting turnstile in the building gate is still required for people's working lives, in general. However, under COVID-19 pandemic, a new norm or New-Normal has emerged to reduce and prevent the spread of the COVID-19 virus. Sound identification system is considered as a system of identification/authentication without any direct contact between the people and the system equipment. Therefore, in this work, a sound identification system is studied and developed. To analyze and feature-extract the sound from a pre-processed human voice, MFCC (Mel Frequency Cepstral Coefficient) technique is adopted. For identification process, the feature vector obtained from MFCC is sent to 3 different popular machine learning techniques.; namely, CNN (Convolutional Neural Network), GMM (Gaussian Mixture Models), and SVM (Support Vector Machine). This results in sound authentication with true positive accuracy of 87.90%, 52.98%, and 41.37%, respectively, and true negative accuracy of 52.98%, 35.12%, and 90.48%, respectively. The best true positive and true negative accuracies are from CNN and SVM, respectively. The results can be further applied in sound identification system.