Comparing Two Models for Word Boundary Detection in a Phoneme Sequence Using Recurrent Fuzzy Neural Networks

Abstract

The word boundary detection has an application in speech processing systems. The problem this paper tries to solve is to separate words of a sequence of phonemes where there is no delimiter between phonemes. This paper tries to compare two conceptual models for word boundary detection, named preorder and postorder models. In this paper, at first, a recurrent fuzzy neural network (RFNN) together with its relevant structure is proposed and learning algorithm is presented. Next, this RFNN is used to post-detecting word boundaries. Some experiments have already been implemented to determine complete structure of RFNN in both models. Here in this paper, three methods are proposed to encode input phoneme and their performance have been evaluated. Some experiments have been conducted to determine required number of fuzzy rules and then performance of RFNN in postdetecting word boundaries is tested in both models. Preorder model shows better performance than postorder model in experiments.