On the Impact of Gabor Phase for Spectro-Temporal Feature Extraction in Building an ASR System

Abstract

Spectro-temporal features have recently showed its usefulness for various speech recognition tasks. Two Dimensional (2-D) Gabor filters are characterized by local patches of spectro-temporal fields which allows them to extract the spectro-temporal information from speech samples. The state of art spectro-temporal features uses the real part of the Gabor filters without any phase offset value. In this work, we analyzed to see whether the Gabor phase have any relevance in the context of spectro-temporal feature extraction in building the acoustic module of a hybrid Automatic Speech Recognition (ASR) system. Different phase offset values are investigated to see if it carries equivalent or complementary information. The experiments are carried out using TIMIT dataset corrupted with different noises at various SNR values. It is found that a Gabor offset phase of 0 degree and 90 degree is equally important in the Gabor filter design for building a robust ASR system.