A fast learning algorithm for adaptive wavelets with application to fuzzy neural-based speaker verification.

摘要

Learning can be observed as a mapping from an input space to an output space. We propose a new learning algorithm, based on Quasi-Newton methods and the Delta learning rule (QND), to extract discriminative wavelet parameters from adaptive wavelet networks for speaker verification. Wavelets are an effective tool in speech signal processing. Adaptive wavelets consist of a weighted linear combination of translated and dilated mother wavelets. Wavelet parameters are adjusted by approximating speech signals adaptively. The objective of learning is to minimize the difference between the original and approximated signals by tuning wavelet parameters adaptively. Quasi-Newton methods are used to adjust dilation and translation parameters in the hidden layer. Coefficients between the hidden and output layer of adaptive wavelets are tuned by the Delta learning rule: The proposed algorithm shows better convergence than a conjugate gradient algorithm in speech signal approximation.;Adaptive wavelets are used to extract a number of wavelet parameters from very short periods of voiced sound. These parameters, which are extracted from specific phonemes, have the properties of low intra-speaker variation, and, at the same time, high inter-speaker variation. These parameters are used as input feature vectors to fuzzy neural networks, which act as a classifier to determine whether the utterance is made by the authorized speaker. The system derives valuable information from each model parameter of each utterance spoken by several speakers to construct a fuzzy rule-based verification system. Membership functions, fuzzy rules, and an inference mechanism are prepared at the training stage. Two different types of data sets are used to evaluate selected features.;Several experiment results will be shown comparing neural networks and fuzzy neural networks in terms of two different types of errors. The fuzzy neural networks reject impostors more accurately that the neural network versions do.