Discrimination Effectiveness of Speech Cepstral Features

In this work, the discrimination capabilities of speech cepstra for text and speaker related information are investigated. For this purpose, Bhattacharya distance metric is used as the measure of discrimination. The scope of the study covers static and dynamic cepstra derived using the linear prediction analysis (LPCC) as well as mel-frequency analysis (MFCC). The investigations also include the assessment of the linear prediction-based mel-frequency cepstral coefficients (LP-MFCC) as an alternative speech feature type. It is shown experimentally that whilst contaminations in speech unfavourably affect the performance of all types of cepstra, the effects are more severe in the case of MFCC. Furthermore, it is shown that with a combination of static and dynamic features, LP-based mel-frequency cepstra (LP-MFCC) exhibit the best discrimination capabilities in almost all experimental cases.