Spectro-temporal Gabor features as a front end for automatic speech recognition

A novel type of feature extraction is introduced to be used as a front end for automatic speech recognition (ASR). Two-dimensional Gabor filter functions are applied to a spectro-temporal representation formed by columns of primary feature vectors. The filter shape is motivated by recent findings in neurophysiology and psychoacoustics which revealed sensitivity towards complex spectro-temporal modulation patterns. Supervised data-driven parameter selection yields qualitatively different feature sets depending on the corpus and the target labels. ASR experiments on the Aurora dataset show the benefit of the proposed Gabor features, especially in combination with other feature streams.