Robustness of Speech Recognition Using Non-Linear Asymmetric Filter and Delta Spectral Characteristics

In this paper, we propose a new feature extraction algorithm which is robust against noise. In the proposed algorithm, a non-linear filter with temporal masking are used for speech feature extraction and by applying delta spectral characteristics instead of delta cepstral, the accuracy of speech recognition is improved. Almost, all present Automatic Speech Recognition (ASR) systems use cepstral-delta and delta-delta characteristics for speech feature extraction. The aim of this paper is to reach the robust speech features which provide more accurate speech recognition under different noisy conditions. This is achieved by focusing on speech key features (especially non-stationary speech features) which highly differ from the noise signals. The obtaining experimental results show that the accuracy of speech recognition improves in comparison with traditional methods such as PLP and MFCC.