Despite its small size, Underdog's new model can generate both speech and music.