Novel Feature Extraction Algorithm using DWT and Temporal Statistical Techniques for Word Dependent Speaker’s Recognition

Abstract

In the paper, a novel method for word-dependent speaker recognition is proposed, based on unique temporal statistical techniques on Discrete Wavelet Transform (DWT) coefficients. The speaker is to be recognized based on words from a specific dataset. For demonstration, the dataset of words used are digits ranging from zero to nine. In presented algorithm, wavelet analysis is being considered as Discrete Wavelet Transform is able to analyze time-frequency multi-resolution for quasi stationary speech signals. The speech signal of a word from dataset, is decomposed using Symlet 7 wavelet as mother wavelet. A 1D feature vector is extracted from approximate coefficients of DWT. Temporal-statistical methods are used to construct unique features for each word in dataset. As compare to methods like LPC, LPCC, MFCC and Power spectral analysis (FFT), the proposed method gives more a better robust feature for speaker recognition. The details of methodology is presented followed by results obtained and discussion.