Results are reported of experiments in the use of Charade, an Automatic Learning System, to classify phonetic macro-classes. A preliminary evaluation of the Charade system is carried out on a reference database of continuous speech and compared to an usual classifier (i.e., Hamming Distance Nearest Neighbor) and a neural net based technique (i.e., Modified Hopfied Net). Preliminary results of classification of phonetic macro-classes can be summarized as follows : For a given reasonable error rate, Charade classifier gives the lowest rejection rate. An important advantage of Charade lies in the ability to analyse and interpret the production rules. For instance, a rule can be interpreted in terms of cues relevant to features.

Spectral masks of vowel, drawn from analysis of the most frequent clustering rules found for each of them, are shown to be coherent with phonetic knowledge.