Cepstral analysis of speech signals in the process of automatic pathological voice assessment

Abstract

The paper describes the problem of cepstral speech analysis in the process of automated voicedisorder probability estimation. The author proposes to derive two of the most diagnosticallysignificant voice features: quality of harmonic structure and degree of subharmonic from cepstrumof speech signal. Traditionally, these attributes are estimated auricularly or by spectrum (orspectrogram) observation, hence this analysis often lacks accuracy and objectivity. The introducedparameters were calculated for the recordings from Disordered Voice Database (Kay, model 4337version 2.7.0) which consists of 710 voice samples (657 pathological, 53 healthy) recorded in thelaboratory environment and described with diagnosis and a number of additional attributes (suchas age, sex, nationality).The proposed cepstral voice features were compared to similar voice parameters derived fromMultidimensional Voice Program (Kay, model 5105 version 2.7.0) in respect to their diagnosticsignificance and presented graphically. The results show that cepstral features are more correlatedwith decision and better discriminate clusters of healthy and disordered voices. Additionally, bothparameters are obtained by single cepstral transform and do not require to perform F0 trackingearlier as it is derived simultaneously.