It wouldn’t have to be a universal voice, just something consistent in the noise. Neat thing with lots of AI training is we don’t have to necessarily even have a theory of what the AI is picking out. It just has to be in the training data and our hints during training have to be appropriate.