ABX-discriminability measures and applications.

Summary This thesis is, at first, an indirect contribution to the problem of modeling the acquisition of phonetic categories in children. The computational models already proposed have not yet been systematically tested to determine whether they are really able to account for a substantial part of the available empirical evidence. We develop an approach allowing a systematic evaluation of models based on ABX Discriminability Measures. We show the interest of our approach by applying it to two related problems: the processing of phonetic categories at birth and in adulthood. The next step will of course be to apply our approach to phonetic category acquisition patterns.The interest of ABX Discriminability Measures is not restricted to the particular case of evaluating phonetic category processing patterns. They are useful in the study of signals other than speech and of categories other than phonetic categories, as well as in disciplinary fields other than cognitive sciences, such as engineering, data mining or artificial intelligence for example. We justify this by studying the properties of these measures in a general abstract framework and by presenting three main families of applications: the evaluation of the ability of systems operating in the absence of explicit supervision to represent a categorical structure. the formulation of simple computational models of behavior in discrimination tasks. the definition of descriptive measures for representations associated with categorical data.