Max-pooling loss training of long short-term memory networks for small-footprint keyword spotting.