Analiza cech suprasegmentalnych jezyka polskiego na potrzeby technologii mowy (review)

This work is a contribution to the study of Polish suprasegmentals, with special emphasis on their classification, modeling, synthesis, and automatic recognition. Combining recent research on speech synthesis with her own research on Polish suprasegmentals, Demenko proposes a model of Polish intonation and then verifies it through speech synthesis.

The book begins with a general introduction (Chs. 1–2) and a brief introduction to the physiology of speech (Ch. 3) followed by the main part of the book (Chs. 4–15). The fifteen chapters are followed by an ‘Appendix’, which provides the texts used to analyze phrase boundaries, the utterances used for perception studies, the algorithms used for backpropagation, and the tabulated results of experiments conducted by D. It closes with a bibliography and an English summary.

Chs. 4–15 can be divided into two parts. Part 1 begins with Ch. 4, a general overview of factors that mark suprasegmental contrasts and phrase boundaries as well as acoustic correlates of accent and suprasegmentals in spontaneous speech. Ch. 5 deals with the problems of linearity and superposition in modeling F0 and F0 declination. Chs. 6–8 return to accent and intonation, focusing on Polish. Ch. 9 looks at vowel duration and intensity in an intonation phrase. This part ends with a discussion of suprasegmentals in continuous speech (Ch. 10). D bases her analysis on a number of speech perception experiments. She not only identifies the features of Polish manifestations of accent but also offers the first explicit model of the Polish intonational phrase.

Part 2, devoted to synthesis research, begins with Ch. 11, a succinct presentation of methods of pitch extraction, algorithms used for pitch determination, and statistical techniques for speech analysis, classification, and recognition. Ch. 12 discusses both the application of neural network techniques to the classification and recognition of intonation patterns and of neural network training to the discrimination of intonation patterns. The following two chapters apply the net architecture and the training process to Polish intonational structures in isolated utterances (Ch. 13) and in connected speech (Ch. 14). D classifies nine nuclear tunes in isolated utterances and tests their automatic classification in different types of neural networks: probabilistic, radial-function, and multilayer perceptron. For connected speech, D classifies five tunes and then attempts to detect accents with a neural network trained on isolated polysyllabic utterances. Ch. 15 presents current (and future) applications of the research in such areas as audiology, speech therapy, voice pathology, and voice organ rehabilitation, giving examples from her own research on pitch perturbation among patients born with hearing disorders, among patients with hearing and speech disorders, and among patients who lost their hearing after age 15.

Focusing more on speech synthesis than on speech analysis, D’s book presents ground-breaking research on Polish suprasegmentals both in terms of its findings and the methodologies employed.