Fuzzy models for recognition of spoken vowels in Romanian language

In the paper we address the problem of vowel recognition in the Romanian language using fuzzy models. The vowels to be recognized are "a", "e", "i", "o", "u" represented by r dimensional feature-vectors with r=3, 6, 10, uttered in a large variety of contexts (over 800) by five male speakers. The ISODATA and k-NN algorithms in hard and fuzzy formulation are applied, the error rate being evaluated for a test set of N/sub T/=500 vowels, for variable size training sets and for three different sizes of the feature vector. The recognition confidence is improved in the fuzzy case versus the hard case, while increasing the dimension of the feature vector.