I think you are wrong to dismiss domain experts opinions as leaving science. Machine learning 101 is eye ball your data, as the human brain is the most powerful pattern matcher known. A musicians ear is therefore the best tool for detecting anomolies in sound and should not be dismissed for not having a nice mathematical construction. Acoustics are perceptual and are heavily filtering the real world sound waves so I don't think its possible to reduce the laudatory cortex to simple filters like its ignoring phase or whatnot.
Want an A/B test? Put your hi-fi in the same room with a real singer, or an acoustic guitar, or whatever musical instrument, and see how hard it is to tell them apart. Not very, I assure you.
Interestingly, we can pick up very subtle and useful musical cues from very poor recordings and reproduction. It's a complex thing.
