Multiple-Instance, Cascaded Classification for Keyword Spotting in Narrow-Band Audio