A syntax-controlled segmentation of speech signal on the basis of dynamic spectra

In this paper an effective algorithm of segmentation of the dynamic spectra of speech signal is presented. The algorithm realizes segmentation into simple elements, very similar to the classic "phonetic-acoustic segments", which contain the full linguistic information carried by the speech signal. The described algorithm is based on the syntactic theory of pattern recognition and makes use of the similarity relations among the momentary spectra composing the dynamic spectrum of an utterance. The proposed method has been experimentally verified, the final correctness of segmentation exceeding 90%.