THE APPLICATION OF ADAPTIVE CONSTRAINT SATISFACTION NETWORKS TO ACOUSTIC PHONETIC ATTRIBUTE DETERMINATION

Speech is a signal with a complex underlying structure and considerable variability. In order to determine acoustic phonetic correlates of speech one must take this structure into account. To specify its structure adequately a priori would be very difficult. One would also have to ensure that any fixed structure imposed initially was not restrictive. Characteristics of learning machines appear useful for this type of problem because they have potential for acquiring internal structure, so less needs to be imposed in advance. The type of learning machine investigated is the multi-layer perceptron (MLP). It is shown that one may train such a system to perform very well at standard pattern recognition tasks. It is compared against a standard technique for discriminant analysis, a Bayes classifier for normal patterns.