Classification of normal and pathological voices using TEO phase and Mel cepstral features

In this paper, a new feature-set, viz., Teager Energy Operator (TEO) phase has been proposed for automatic classification of normal vs. pathological voices. Development of TEO phase has been motivated from recently proposed linear prediction (LP) residual phase for speaker recognition. Classification was performed using a discriminatively-trained 2^nd order polynomial classifier on a subset of the Massachusetts Ear and Eye Infirmary (MEEI) database. Score-level fusion of TEO phase and state-of-the-art Mel frequency cepstral coefficients (MFCC) gave reduction in equal error rate (EER) by 1.86 % than EER of MFCC alone. Proposed TEO phase feature set is also evaluated under degraded conditions using the NOISEX-92 database for the case of additive car noise.