Comparative analysis of filipino-based rhinolalia aperta speech using mel frequency cepstral analysis and Perceptual Linear Prediction

Abstract

In this work, a database collected from two different Filipino Cleft Palate patients was used to identify the discriminative features for hypernasal speech. Data from the Filipino Speech Corpus (FSC) were used as normal speech samples. The features identified were based from three feature extraction algorithms, Mel Frequency Cepstrum Coefficient (MFCC), Perceptual Linear Prediction (PLP), along with a MFCC-PLP hybrid feature extraction method, which was introduced in this study. Intraclass and interclass correlation among the speech samples, separating the two hypernasal speech samples and along with the normal speech samples were computed to determine the correlation of the speech samples to each other. This paper will also compare the differences between the extracted MFCC features, PLP features and a hybrid of MFCC and PLP features to determine the most discriminative features from hypernasal speech compared with normal speech and the most discriminative features from hypernasal speech obtained from different study volunteers through Analysis of Variance (ANOVA) statistical analysis. The p-values obtained from the ANOVA test will be the basis to determine which features provide a certain degree of significant difference between speech samples. The paper will also present and determine the most optimal and conclusive feature extraction method in analyzing speech samples using MATLAB and through correlation analysis.