Vocal Features for Glottal Pathology Detection using BPNN

Development of low cost, non-invasive applications is one of the most challenging tasks in the field of biomedical signal processing. Present work focuses on detection of glottal pathology with the knowledge of prominent speech processing and machine learning techniques. This paper addresses the discriminative characteristics of speech signal like, pitch, jitter, linear prediction residual and cepstral source excitation to aid such an identification system. Back-propagation Neural Network model is developed for various feature combinations to classify the glottal pathologic voice from normal voice. Accuracy of the developed system is evaluated considering different feature sets. Work also concludes the efficiency of such acoustic features for various combinations using objective measures like confusion matrix, true positive rate i. e. sensitivity, specificity i. e. true negative rate and accuracy. The results show promising development in identification of pathological individual from normal person using voice samples.