Feature space reduction and classification in automatic voice quality estimation

The paper presents an approach to solve a problem of feature space reduction applied to a voice quality estimation system. In order to reduce symptom space dimensionality, a method based on of decision tree and Fisher Linear Discriminant has been introduced. On the basis of 3 discrimination tests, the patient's voice is being classified into one of 3 groups - healthy, ill or risk, obtaining 90% of correct results. The experiment involved voice recordings of 70 patients who were diagnosed by the specialist. The method has been applied to a system of automatic voice quality estimation - the SpeechAnalyser, which was designed to be a supportive tool in laryngologica. l screening tests and treatment progress monitoring. There have been also briefly introduced the algorithms of feature extraction from a voice sample and also diagnostic significance of the symptoms has been discussed. Author proposed a new approach to cepstral analysis that allows objective measuring of harmonic and subharmonic content is spectrum.