Clustering analysis of phonetic and text feature vectors

Abstract

Our goal is to show an example of using statistical methods to analyse some attributes of speeches. For this purpose, the New Year’s Day speeches of Czech and Czechoslovak presidents are chosen. The aim of our study is researching similarities among these speeches and their recognizability through the history of Czechoslovak politics. All presidents are compared between each other. The comparison method is based on principal component analysis and cluster analysis. Important part is creating a feature vector. The feature vector doesn't have to be the same for successful clustering. There are many varieties and combinations of features that can be selected and used. Correlated variables must be discarded. The most significant features are chosen to represent and characterize the speaker. Some speakers can have something in common according to the chosen features. Or on the other hand they can differ much more from others. This kind of approach can help us to recognize a speech pattern of each spokesman independently.