Sparsification via Compressed Sensing for Automatic Speech Recognition