Classification of stuttering events using I-vector

Abstract EN

Stuttering represents the main speech disfluency problem with the most two common stuttering disfluencies events are repetitions and prolongations.

It is most desired to classify these disfluencies automatically rather than manually classification, which is a subjective, time-consuming task, and depends on speech language pathologists experience.

In the proposed work, a new automatic classification approach is presented which depends on using the i-vector methodology that was usually used only in speaker verification/recognition applications, a sufficient accuracy relative to the amount of data used resulted as 52.43% ,69.56%,40%,50% for normal, repetition, prolongation, rep-pro1 classes respectively and 64.75%,71.63% for normal, disfluent classes.

Best accuracies for classifying the rep.

and pro.

classes with equal number of samples in each class resulted from the ivector approach with 77.5%, 82.5% for rep., pro respectively compared to the Mel-Frequency Cepstrum Coefficients/Linear Prediction Cepstrum Coefficients (MFCC/LPCC)- K-Nearest Neighbour/Linear Discriminant Analysis (KNN/LDA) approaches tested on the same data set.