A Neural Network Based Approach to Objective Voice Quality Assessment

Abstract

Voice quality is of fundamental importance to the patient following treatment of cancer of the larynx. Current techniques for voice analysis are slow, mainly subjective, and based on limited numbers of retrospective studies. This study is concerned with the development of an on-line system which encapsulates the expert knowledge of the Speech and Language Therapist in such a way as to provide an objective and consistent assessment of voice quality for staging and treatment monitoring of cancer of the larynx.

After discussions with the Speech and Language Therapist it was concluded that their expert knowledge was related to subtle variations in the frequency structure in a patient’s stylised speech. In order to identify the frequency components that can be used to provide an objective classification and assessment of a patient’s voice quality, appropriate parameters were extracted from a segment of speech recorded from 20 male patients with cancer of the larynx and 20 male volunteers who were considered as having normal voice quality. These parameters were then presented to a feedforward Artificial Neural Network known as the Multi-Layer Perceptron. This Multi-Layer Perceptron was shown to be able to distinguish between normal, i.e. non-cancerous, subjects and patients having cancer of the larynx, achieving a classification accuracy of between 85% and 90%. These results provide the basis for an extension of this work into a practical system that may be utilised by the Speech and Language Therapist during clinical examinations to provide an objective measure of voice quality.