SPEECH NONFLUENCY DETECTION AND CLASSIFICATION BASED ON LINEAR PREDICTION COEFFICIENTS AND NEURAL NETWORKS
Adam Kobus, Wiesława Kuniszyk–Jóźkowiak, Elżbieta Smółka, Ireneusz Codello · Journal of Medical Informatics & Technologies · 2010
The goal of the paper is to present a speech nonflu ency detection method based on linear prediction co efficients obtained by using the covariance method. The applic ation “Dabar” was created for research. It implemen ts three different methods of LP with the ability to send co efficients computed by them into the input of Kohon en networks. Neural networks were used to classify utterances in categories of fluent and nonfluent. The first one was Kohonen network (SOM), used to reduce LP coefficients representation of each window, which were used as input data to SOM input layer, to a vector of winning neurons of SOM output layer. Radial Basis Function (RBF) networks, linear networks and Multi-Layer Perceptrons were used as classifiers. The research was based on 55 fluent sam ples and 54 samples with blockades on plosives (p, b, d, t, k, g). The examination was finished with the outcome of 76% classifying.