Optimized Efficientnet-B0 Framework for Speech Pathology Classification
Roohum Jegan, Chanchal Sharma, Manjusha Pradeep Deshmukh · 2025
This paper introduces a speech disorder classification framework utilizing optimized EfficientNet-B0 features. The methodology involves converting healthy and pathological speech samples from the SVD dataset into mel-spectrogram images, followed by feature extraction using EfficientNet-B0. The spatiotemporal features extracted from the spectrograms are then optimized using the Marine Predator Algorithm (MPA). MPA optimizes the features by maximizing classification accuracy and minimizing feature dimensions, effectively reducing redundancy. The classification task is performed using a KNN classifier. The effectiveness of the algorithm is assessed based on sensitivity, accuracy, precision, specificity, and the$\mathbf{F}$-measure. The results demonstrate the algorithm's effectiveness in classifying speech samples from the SVD database.