Front-end post-processing using histogram equalization combined with ARMA filtering for noise robust speech recognition

In this paper, we present a new feature set for robust speech recognition based on histogram equalization (HEQ) combined with auto regressive moving average (ARMA) filtering. Cepstral vectors extracted from the clean data, modified by Mean and Variance Normalization (MVN) have been used to generate a reference histogram for histogram equalization. The proposed post-processing module also consists of ARMA temporal filtering applied to normalized cepstral coefficients. HEQ compensates for nonlinear distortions caused by noise and ARMA filtering is used for smoothing the normalized feature vectors. The results on the AURORA2 task have shown noticeable improvements in the recognition of noisy speech. The proposed front-end achieved a relative error reduction of around 60% compared to the standard Mel-Cepstral front-end.