Multi-resolution Speech Recognition Based on Hyper-rectangular Fuzzy System

Abstract

This paper presents a multi-resolution feature extraction technique to speech recognition. The proposed multi-resolution feature extraction technique uses wavelet transform and wavelet packet to calculate features of each sub-band in order not to spread noise distortions over the entire feature space. In our previous works, we had developed a method for speech classification. For speech classification, the universe of discourse is divided into many types, and each type is treated as a class. The hyper-rectangular fuzzy system is used to classify frames and integrate the rule-based approach. The variances of each sub-band are utilized to extract both crisp and fuzzy classification rules. In our experiments, the Texas Instruments/Massachusetts Institute of Technology database is used and extracts features of phonemes. The results demonstrate the superior performance to Mel frequency cepstral coefficients. The effectiveness of the proposed system is encouraging.