Analysis of Frequencies and Pitches for Vocal Cord Paralysis Classification Through Transfer Learning

Audio-based examination for Vocal Cord Paralysis (VCP) is a useful, non-invasive, cost-effective procedure, and Spectrogram analysis is one such boon for the classification of VCP versus Healthy subjects. However, spectrograms have not been utilized fully from their representation for classification purposes. This research aims to investigate spectrograms at various frequencies over different pitches to aid in accurately classifying VCP and healthy subjects. A voice disorders database named Saarbruecken Voice Database (SVD) contains audios of VCP and healthy subjects are collected. Using Transfer Learning (TL) and fine-tuning, Deep Convolutional Neural Networks (DCNNs) and Vision Transformers (ViTs) architectures are trained over the vowel /a/, and the derived characteristics are given for classification, yielding outstanding 100% accuracy. With minimal trainable parameters, the Deep Learning (DL) architectures performed exceptionally well, indicating the computational capability of the spectrograms in terms of robustness for pathological practice in the future.