Normalization of spectro-temporal Gabor filter bank features for improved robust automatic speech recognition systems

Physiologically motivated feature extraction methods based on 2D-Gabor filters have already been used successfully in robust automatic speech recognition (ASR) systems. Recently it was shown that a Mel Frequency Cepstral Coefficients (MFCC) baseline can be improved with physiologically motivated features extracted by a 2D-Gabor filter bank (GBFB). Besides physiologically inspired approaches to improve ASR systems technical ones, such as mean and variance normalization (MVN) or histogram equalization (HEQ), exist which aim to reduce undesired information from the speech representation by normalization. In this study we combine the physiologically inspired GBFB features with MVN and HEQ in comparison to MFCC features. Additionaly, MVN is applied at different stages of MFCC feature extraction in order to evaluate its effect to spectral, temporal or spectro-temporal patterns. We find that MVN/HEQ dramatically improve the robustness of MFCC and GBFB features on the Aurora 2 ASR task. While normalized MFCCs perform best with clean condition training, normalized GBFBs improve the ETSI MFCCs features with multi-condition training by 48%, outperforming the ETSI advanced front-end (AFE). The MVN, which may be interpreted as a normalization of modulation depth works best when applied to spectro-temporal patterns. HEQ was not found to perform better than MVN.