On the Relevance of Auditory-Based Gabor Features for Deep Learning in Automatic Speech Recognition