Two-way voice feature representation for disease detection based on voice using 1D and 2D deep convolution neural network

Abstract

Voice pathology deals with detecting diseases with the help of the voice, as diseases significantly impact the voice. Machine learning (ML) and deep learning (DL) schemes have been presented for disease detection using voice. However, the outcomes of the system are limited due to poor spectro-temporal representation, less feature distinctiveness, low-frequency resolution problems, lower detection rates, etc. This article presents voice-based pathology using two-way voice feature representation (TWVFR), which consists of two parallel arms of a Deep Convolution Neural Network (DCNN) for feature representation. The first parallel arm considers the Mel Frequency Cepstral Coefficient Spectrogram (MFCCS), fed to 2-D DCNN to characterize the spectral domain characteristics of the voice signal. The second approach consists of multiple voice features (MVF), such as spectral domain (SD), time domain (TD), and voice quality (VQ) features. The essential features are selected using the Spider Monkey optimization algorithm and given to 1D-DCNN. The last layer features are combined and given to a fully connected layer followed by the Softmax classifier. The Softmax classifier classifies the speech signal into normal and diseased voices. The system outcomes are validated on the Saarbruecken Voice Dataset (SVD) for four class disease classifications: Bulb Paralysis, Cyste, Polyp, and Normal. The suggested TWVFR scheme provides improved overall accuracy of 98.33%, recall of 0.98, precision of 0.98, F1-score of 0.98, selectivity of 0.98, and negative predictive rate of 0.98 compared to existing methods. The TWVFR helps to enhance the feature depiction and provides an overall accuracy of 98.33% than MVF-1D DCNN (95.45%). The suggested TWVFR scheme provides improved overall accuracy of 98.33%, recall of 0.98, precision of 0.98, F1-score of 0.98, selectivity of 0.98, and negative predictive rate of 0.98 compared to existing methods.