Performance Evaluation of Pre-Trained Models for Classification of Vocal Cord Paralysis over Vowels

This paper intends to explore the use of various Deep Convolutional Neural Networks(DCNNs) and Vision Transformers(ViTs) in the classification of Vocal Cord Paralysis(VCP). A non-invasive approach to classify VCP using spectrograms can be an effective viewpoint for scrutinizing. From the Saarbruecken Voice Database (SVD) database, audio samples of the vowels /a/, / i /, and / u / at various pitches are considered and sampled at different frequencies. Utilizing the Transfer Learning(TL) along with various models, a remarkable accuracy of 93% for ViT-B/16 and 88% for MobileNet is observed for the vowel/a/ at 44.1 KHz. The vowels /u/ and /i/ also performed exceptionally well with an accuracy of 90% at 50 KHz and 87% at 16 KHz, respectively, expressing the distinguishing capability of the models over an imbalanced and limited dataset, which controlled the notion of the need for humongous input. As a result, tuning the frequency and vowels can help automate the analysis of the voice-associated pathology VCP.