What type of signal analysis (if any) might be performed on a recorded music vocal track to determine if pitch correction might have been applied to a performance?
1 Answer
someone recently pointed me to a paper that purports to detect and count duplicated fragments. i presume that if the density of duplicated fragments exceeds some threshold, one might suspect that up-shifting may have been done to that piece of audio. i don't see how it could detect down-shifting (how does it detect a removed fragment?).
if, by "pitch correction", you mean pitch quantizing or autotuning™ on a monophonic sound, then i suppose you might run a good tracking pitch detector (maybe not such an easy problem) and maybe a histogram on the detected pitches. we would expected pitch corrected vocals to have, in the histogram of pitch, a higher frequency of occurrence for pitches more closely located adjacent to pitch semitones (if autotune was set to correct to equal-temperament) or against some note identified as the tonic and then other notes that are "just intoned" (frequency ratios of small integers) with that tonic.
pitch detection/estimation is a signal processing issue, but determining what the tonic is, might be sorta a musicological question. also something could be pitch "corrected" to a pitch standard that is not A-440. perhaps all of the semitone locations are scooted down a quartertone from A-440.
if one is not going for the "Cher effect", pitch correction can be very subtle when it is done well. one need not clamp the pitch hard against the semitone targets. one can cause the amount of pitch correction to "slew" a little, which makes it slower to correct. one can separate the vibrato out of the pitch detector output (with a LPF applied to the pitch contour signal), apply pitch correction to the remaining pitch contour, then add the extracted vibrato back. that would be very hard to detect. maybe the singer just sang it well.
not an easy problem. and not a well-solved problem, at least with the current state-of-the-art.