run voice through 3 models, when one model is different than the other two then throw it away.

problems abound with that approach, but I could envision it.