Adaptive ARMA filtering and energy normalization for robust speech recognition

In this paper we evaluate the effect of applying a previously proposed energy normalization scheme on a simple noise robust feature post processing scheme, MVA, in automatic speech recognition (ASR). A Noise energy-adaptive approach has also been proposed and successfully implemented. Here, the integration of non-logarithmic raw energy normalization in cepstral mean and variance normalization and applying an auto-regression moving-average (ARMA) filtering has led to improvements of up to 58% compared to the Aurora 2 clean-trained baseline system and up to 24% compared to MVA alone. Further improvements are also obtained by applying the adaptive approach. Our method is also very simple to implement and hence can easily be integrated with features other than the standard MFCC such as LPC, PLP etc.