Noise robust features based on MVA post-processing

Abstract

In this paper we present effective technique to improve the performance of the automatic speech recognition (ASR) system. This technique consisting mean subtraction, variance normalization and application of temporal auto regression moving average (ARMA) filtering. This technique is called MVA. We applied MVA as post-processing stage to Mel frequency cespstral coefficients (MFCC) features and Perceptual Linear Prediction (RASTAPLP) features, to improve automatic speech recognition (ASR) system. We evaluate MVA post-processing scheme with aurora 2 database, in presence of various additive noise (subway, babble because, exhibition hall, restaurant, street, airport, train station). Experimental results demonstrate that our method provides substantial improvements in recognition accuracy for speech in the clean training case. We have completed study by comparing MFCC and RSTA-PLP After MVA post processing.