Speaker independent recognition, analysis, and understanding of utterances of spoken language is much more complicated than the classification of simple patterns: with each simple pattern a feature vector of known dimension can be computed, and decision rules introduced in Chapter 24 can be applied to solve the classification task. As it was already shown, speech signals are decomposed into frames. Thus, each speech signal is associated with a sequence of features. The number of sequence elements depends on the duration of the utterance and varies for each signal. Obviously, the simple application of the Bayesian classifier or the nearest neighbor decision rule, as defined for single feature vectors, is not possible. Extensions of these decision procedures are required, even for the implementation of a single word recognition system.
Unable to display preview. Download preview PDF.