A Chinese team's hybrid attention model read acoustic cues from speech and cleared 95 percent on precision, recall and F1.