If you sample the audio from machines with better and better efficiency. I assume its then possible to imagine or predict the sound from a slightly better machine using machine learning. Then with this predicted audio sample. Then I wonder. Could you then use this together with other data to construct this slightly better machine with the target audio?