When spoken words, tone of voice and facial expression disagree, listeners infer feelings mostly from the non-verbal channels (7% words, 38% tone, 55% face).