Automatic Speech Recognition on Non-Pathological Dataset of Urdu Language

Abstract

One of its subsystems, speech, has a strong underlying characteristic and a distinct voice. Voice disorders are abnormal conditions that influence the quality of voice. Several protocols, including acoustic analysis, can detect clinical voice pathology. Based on a computerized acoustic analysis, machine learning algorithms and non-invasive systems may play a very vital part in initial detection, tracking, and even growth of proficient pathological speech analysis. The aim of this research paper is to collect a non-pathological dataset i.e. healthy voice dataset. Two important and critical features; 1) MFCC and 2) Pitch are used to generate a final audio clip. SVM used as a classifier to train and test the dataset model and the models exhibited reasonably high training and testing accuracies i.e. 85.886% which proves to be a milestone on Urdu language dataset.