Abstract

Several clustering techniques were used for finding subgroups of speakers sharing common characteristics within a sample of 14 dysarthric speakers and 15 non-dysarthric speakers. Our classifying variables were five spectral measures computed from the temporal envelope of each of the four sentences read by the participants. The unsupervised k-means clustering algorithm showed that the optimal number of clusters in this dataset is two, with Cluster 1 matching almost exactly the dysarthric population and Cluster 2 the non-dysarthric population. As for the importance of each variable, a PCA analysis revealed that centroid, spread, rolloff and flatness contribute equally to the first component, and entropy contributes to the second component. Hierarchical agglomerative clustering further supported the separation into two main clusters (highlighting the relevance of these rhythmic measures to characterize dysarthria), but also allowed us to detect possible subgroups within each main speaker group.