Machine Learning based vowel analysis in severity level detection of Voice Pathology

A correct classification of voice pathology is essential for pedagogic purposes as well as to aid a clinician in establishing the “primary diagnosis,” i.e., an initial voice issue in an individual patient. Numerous experimental research using quantitative and qualitative analysis to find vocal disorders have been conducted. However, the repeatability of the experimental outcomes in laboratory settings is uncertain. In this study, a method to classify voice recording into healthy or pathological classes is done which will aid in the early diagnosis of underlying speech and pathological defects. Voice data is acquired from the Saarbruecken Database, which comprises various voice clippings collected from different subjects with a wide range of voice pathologies and healthy voice recordings. The Mel-frequency cepstral coefficients (MFCC) features were used to determine the test accuracies of the three vowel sounds /a/, /i/, /u/. This work serves as a comparative study between different types of vowel sounds such as /a/, /i/, and /u/ sounds, and a comparison between Neural Network (NN) and Deep Neural Network (DNN) techniques has also been done to determine the model that is best for the detection of voice pathologies. The DNN model showed an accuracy of 92.3 % for the /u/ sound compared to other vowel sounds.