Arabic Speech Recognition Using Two Techniques Hybrid & 3D-Multiwavelet

Abstract

A key issue for implementing an accurate speech recognition system is the set of acoustic features extracted from speech signal. This paper presents two techniques for comparison. The first technique converts successfully the speech signal from (1-D) into two dimensional (2-D) forms. Next, the 2-D Multiwalidlet transform is applied to each 2-D signal. The second used transformation which is 3D-Multiwavelet (DMWT). For this transform set of speakers spoke the same word which arranged as slices of 2-D signals in acoustic space. These speakers represented the word as 3-D signal. The techniques apply the neural network as a classifier and dealing with text-dependent and text-independent speech recognition. The works are tested upon a database which consist of (28) speakers and uttered 7 Arabic words for each one. It was compared with first technique which gave the result (85.71%-100%), the second gave (71.43%-100%). It is clear that first give much better performance than the second one.