Classification of voiced and non-voiced speech signals using empirical wavelet transform and multi-level local patterns

This paper presents a novel algorithm for classification of voiced and non-voiced speech segments in noisy environment. Empirical wavelet transform (EWT), an adaptive technique for analyzing non-stationary signals, is employed in the pre-processing stage for suppression of noise in speech signals. In this work, multi-level local patterns (MLP), modified version of 1D-local binary patterns (LBP) are used as features. Multi-level local patterns capture the local variations in non-stationary signal by performing comparisons in neighborhood of a sample. Finally, the comparative information thus generated is encoded into multiple states and histogram of MLPs corresponding to short segments of speech signal is computed. Nearest neighbor classifier utilizes the histogram features for classification of speech segments. Experimental evaluation of proposed approach is carried out on the publicly available CMU-Arctic database. The results of our experiments show improvement in classification accuracy with the use of EWT. Further, the MLP based approach clearly yields superior performance than the LBP based approach.