Search

This book is devoted to pattern analysis, that is, the automatic construc­ tion of a symbolic description for a complex pattern, like an image or con­ nected speech. Pattern analysis thus tries to simulate certain capabilities which go without saying in any human central nervous system. The increasing interest and growing efforts at solving the problems related with pattern analysis are motivated by the challenge of the problem and the expected ap­ plications. Potential applications are numerous and result from the fact that data can be gathered and stored by modern devices in ever increasing extent, thus making the finding of particular interesting facts or events in these hosts of data an ever increasing problem. It was tried to organize the book around one particular view of pattern analysis: the view that pattern analysis requires an appropriate set of modules operating on a common data base which contains interme­ processing diate results of processing. Although other views are certainly possible, this one was adopted because the author feels that it is a useful idea, be­ cause the size of this book had to be kept within reasonable bounds, and because it facilitated the composition of fairly self-contained chapters.

This volume contains invited and contributed papers presented at the NATO Advanced study Insti tute on'Recent Advances in Speech Understanding and Dialog systems'held in Bad Windsheim, Federal Republic of Germany, July 5 to July 18, 1987. It is divided into the three parts Speech coding and Segmentation, Word Recognition, and Linguistic Processing. Although this can only be a rough organization showing some overlap, the editors felt that it most naturally represents the bottom-up strategy of speech understanding and, therefore, should be useful for the reader. Part 1, SPEECH CODING AND SEGMENTATION, contains 4 invited and 14 contributed papers. The first invited paper summarizes basic properties of speech signals, reviews coding schemes, and describes a particular solution which guarantees high speech quality at low data rates. The second and third invited papers are concerned with acoustic-phonetic decoding. Techniques to integrate knowledge­ sources into speech recognition systems are presented and demonstrated by experimental systems. The fourth invited paper gives an overview of approaches for using prosodic knowledge in automatic speech recogni tion systems, and a method for assigning a stress score to every syllable in an utterance of German speech is reported in a contributed paper. A set of contributed papers treats the problem of automatic segmentation, and several authors successfully apply knowledge-based methods for interpreting speech signals and spectrograms. The last three papers investigate phonetic models, Markov models and fuzzy quantization techniques and provide a transi tion to Part 2.