Machine Learning based human recognition via robust Features from audio signals

Abstract

Biometric verification techniques are commissioned throughout the world in different applications. Human voice recognition is one of the biometric techniques. This technique consists of identifying a human from their voice characteristic. This popular and beneficial biometric technique could be employed for identity human, security purposes, and many different applications. Human audio signal recognition consists of two phases i.e., Features Extraction and Classification. The proposed work consists of extracting features through the Mel Frequency Cepstral-Coefficient (MFCC) from the human audio signal, selecting robust features through Principal Component Analysis PCA, and classifying the selected features by comparing seven Machine Learning and proposed deep learning algorithms. Finally, compare the performance of different algorithms with different percentages of selected features to evaluate the acceptance rate of the correlated features. Support Vector Machine SVM shows the best performance with an Accuracy of 99.27% with F1-score and ROC values of 1.00. In the comparison with other methods, the Random Forest and CNN-ANN are 2ndtop robust models with an accuracy of 98.7%. Some of the algorithm's accuracy decreased with fewer features, including Naïve Bayes accuracy suddenly decreases to 60% on 20% of total features. The experiment concludes the acceptance rate of correlated features in different ML and DL algorithms are different in speech processing data.