Fuzzification of formant trajectories for classification of CV utterances using neural network models

摘要

In this paper we show that fuzzification of formant data of a sequence of frames in the transition region of a CV utterance improves recognition of CV utterances. Reliable spotting of CV segments in continuous speech can significantly improve the performance of a speech-to text system. Formant transitions in the transition region of a CV segment provide important clues for recognition of stop consonant CV segments. Therefore, it is necessary to obtain a suitable parametric representation of speech data in the transition region of a CV segment to be used as input to a classifier. We discuss the choice of formants as features representing the CV segments and the fuzzy nature of these features. The details of a fuzzy neural network classifier based on the ideas given by Pal-Mitra (1992) are discussed. Methods for fuzzification of formant trajectories are presented. Results of studies on recognition of CV segments using different methods of fuzzification of formant data are given.