Lessons from Building Acoustic Models with a Million Hours of Speech