Neural Network based Detector for Voice Pathology using Critical band Energy Spectrum

Shama, Kumara and Ananthakrishna, T and Cholayya, Niranjan U. (2005) Neural Network based Detector for Voice Pathology using Critical band Energy Spectrum. In: Proceedings of IFBME 12th International Conference on Biomedical Engineering (ICBME 2005), 7 to 10 December 2005, Singapore.

[img] PDF
conf59a279.pdf - Published Version
Restricted to Registered users only

Download (119kB) | Request a copy


Acoustic analysis is a useful non-invasive technique for the detection and diagnosis of voice pathology. In this paper we demonstrate a neural network based classifier for the automatic detection of voice pathology. Multi-layer perceptron network architecture is designed. Pathologic voices contain additional noise components due to the malfunctioning of vocal folds and these noise components bring significant changes in the energy spectrum of voiced speech. The normalised energy at critical bands is effectively used to differentiate the pathologic voices from normal voices. The critical bands have centre frequency and bandwidths that roughly correspond to human auditory neurons and hence the proposed system could be used as an supplement to the perceptual evaluation of voice for pathology detection. The classifier is tested on voice samples taken from a database and a frame accuracy of 93.52% has been reported. We have also tested the classifier with different number of critical bands, but best results were obtained with 21 bands.

Item Type: Conference or Workshop Item (Paper)
Subjects: Engineering > MIT Manipal > Electronics and Communication
Depositing User: Anantha krishna T
Date Deposited: 11 Jul 2011 05:29
Last Modified: 25 Feb 2015 05:12

Actions (login required)

View Item View Item