The application of Kohonen and Multilayer Perceptron Networks in the speech nonfluency analysis

Paper reports the neural network tests on ability of recognition and categorising the nonfluent
and fluent utterance records. 40 of 4-second fragments containing the blockade before
words starting with stop consonants (p, b, t, d, k and g) and including from 1 to 11 stop consonant
repetitions and 40 recordings of the speech of the fluent speakers containing the same
fragments were applied. Two various networks were examined. The first, Self Organizing
Map (Kohonen network), with 21 inputs and 25 neurons in output layer, was used to reduce
the dimension describing the input signals. As a result of the analysis we achieved vectors
consisting of the neurons winning in a particular time point. Those vectors were taken as an
input for the next network that was Multilayer Perceptron. Its various types: with one and two
hidden layers, different kinds and time of learning were examined.