Recognition of repetitions using Support Vector Machines

The goal of this paper is to present experimental results for the automatic recognition of dysfluencies in the stuttered speech. Mel Frequency Cepstral Coeficients reduce the dimensionality of data and models of acoustic waves of human speech. The acoustic model contains the feature vectors of speech used for further processing with Support Vector Machine. SVM classifier with kernel functions efficiently carries out computations in higher dimensions. Our results compare SVM classifier efficiency with multimodal kernel functions. For the group of 16 speakers who stutter, the SVM classifier with unimodal kernel functions recognizes fluent and dysfluent segments in speech with 96.133% accuracy, while the SVM classifier with multimodal kernel functions reached the 96.4% accuracy.