Robust Speech Recognition Based on Mixed Histogram Transform and Asymmetric Noise Suppression

This paper proposes a new feature extraction algorithm which is robust against noise using histogram compensation and asymmetric filter. Temporal masking is used to improve Automatic Speech Recognition (ASR) systems specifically in matched and multi-style training conditions. Nonlinear filtering and temporal masking are used in the proposed algorithm. By matching the power histograms of the input in each frequency band to those obtained over clean training data, and then mixing the processed and unprocessed spectrums together, speech recognition accuracy can be appropriately increased. The obtained results show that recognition accuracy in comparison with Mel Frequency Cepstral Coefficients (MFCC), Perceptual Linear Prediction (PLP) and Power Normalized Cepstral Coefficients (PNCC), improves in various training conditions and different SNRs.