Max-Pooling Loss Training of Long Short-Term Memory Networks for Small-Footprint Keyword Spotting