Speech recognition system based on short-term cepstral parameters, feature reduction method and Artificial Neural Networks
Nawel Souissi, Adnane Chérif · 2016
The acoustic analysis can provide great results in the identification of voice disorders as a complementary tool to other medical techniques. This paper scrutinizes the Mel Frequency Cepstral Coefficients (MFCC), their first and second derivatives. A full comparative study is established in order to demonstrate that short-term cepstral parameters could be useful to conclude an efficient system for detecting voice impairments. In this context, a projection based Linear Discriminant Analysis (LDA) is investigated to improve the discriminatory ability of the proposed system. In addition, every feature combination is classified by Artificial Neural Networks (ANN). Moreover, the system performance is evaluated in terms of accuracy, sensitivity, specificity, precision and Area Under Curve (AUC). Finally, our findings demonstrate that the optimized combination of the original MFCC features with their first and second derivatives provides the best performances with an accuracy rate of 87.82% and AUC of 87.96%.