Speech syllabic structure extraction with application of Kohonen network

Kohonen network was applied in the analysis of fluent and non-fluent utterances of any length. The input vector consisted of sound level values measured in dB(A) in 1/3-octave bands. Fragments of stuttering people’s and fluent speakers’ utterances were analysed. The optimum neural network was chosen of a 5×5 rectangle topology. Time arrangement of winner neurons in the network extracts single disfluencies, or, in the case of fluent courses, reflects the syllabic structure of an utterance.