Enhancing Voice Disorder Detection Using Deep Transfer Learning Feature Fusion

This paper introduces a computerized non-invasive voice pathology detection system using deep transfer learning network (DTLN) feature fusion. The system takes both healthy and pathological voice samples as input and converts them into mel-spectrogram visual representations. Subsequently, it employs three deep learning architectures, namely (a) AlexNet, (b) ResNet-50, and (c) Inception-V3, to extract complex features from the input signal's spectrograms. As the feature vector dimensions grow due to the aggregation of features from these three CNN models, the study employs an infinite feature selection algorithm to identify the most distinguishing features. These selected optimal features are then used to classify input speech samples as either healthy or pathological, utilizing a K-nearest neighbor (KNN) classifier. The effectiveness of this method is evaluated on three well-established speech pathology datasets, namely AVPD, SVD, and PdA, using metrics such as precision, specificity, sensitivity, F-measure, and accuracy. The experimental results reveal that the proposed deep feature fusion approach achieves accuracy rates of 97.86%, 95%, and 96.83% for the SVD, AVPD, and PdA datasets, respectively.