Except it will a long way down. The random noise would have to match the characteristics of human speech for there to not be information that's possible to gather in an uncoded system.
Actually hmm, I don't know if machine learning or neural networks or genetic programming have been tried on this, but it sure does sound like something they could work at!