Stuttering Classification by KNN Model: Study of Dimension Reduction, Distance Metrics and Number of Neighbors

Stuttering is a speech disorder that disrupts the normal flow of speech. Automatic stuttering classification has got the attention of many researchers. The $\mathbf{k}$-nearest neighbor (KNN) algorithm is a simple, non-parametric and supervised machine learning model used for data classification. It has only two parameters that can be varied to seek reasonable classification accuracy. This paper presents an inclusive study on KNN model for stuttering classification. In this study, a set of examples of three classes of stuttering, normal, sound repetition and prolongation corpuses, has been extracted from the Kassel State of Fluency (KSoF) database. The speech corpus examples selected are the ones having unique labels in the dataset. A set of acoustic features including MFCCs, their delta and delta-delta in addition to the number of zero-crossings and the spectral frequency centroid are selected experimentally for training and testing the KNN model. The average and standard deviation of the classification accuracy are computed from 50 Monto Carlo Runs against the size of the dimensional-reduced features, obtained by the principal components analysis (PCA) and the factor analysis (FA); the number of neighbors k and the distance metrics for the KNN model. Computer results have shown that for dimensional reduction, PCA provides higher accuracy than FA. Results also show that 30 PCs for dimensional reduction, 5 neighbors and the Minkowski distance metric provide accuracy of $0.9635 \pm 0.0176$ in the classification of the three classes.