Choosing Optimal Value for Fuzzy Membership in FCM Algorithm for LP-Residual Input Features

The state-of-art speaker recognition system employs vocal tract information for modeling through different supervised and unsupervised models. Whereas, the baseline of the paper uses LP-residual as the acoustic feature for the following studies. Fuzzy C-Means (FCM) is used to model the information extracted from LP-residual to develop speaker models. FCM is a well-known unsupervised fuzzy model used in speech recognition. Speakers are modeled in order to develop a text-dependent Automatic Speaker Verification (ASV) system. The performance of FCM model have been observed for different codebook sizes varying from 32 to 1024. Also studies are carried out for different fuzzy membership values varying from 1.39 to 4. For LP-residual features, the performance of FCM remains unchanged even after changing the codebook sizes whereas with the change of fuzzy membership the performance of the system is observed to vary.