Octave Errors in Automatic Pitch Tracking
The characteristic failure of autocorrelation: a signal also lines up with itself at two periods or at half a period, so the estimate lands an octave low or high — a factor-of-two error in a phenotype. Search-range settings chosen for one kind of voice and applied to another are the usual trigger, which is a direct hazard for any study pooling voices across a ~100 Hz sex difference in pitch.
Questions this Concept answers
- Why does an autocorrelation-based tracker produce errors of exactly a factor of two?
how to choose a pitch analysis method
Autocorrelation pitch trackers estimate F0 from the time shift at which a waveform best matches itself. Because a signal periodic at 5 ms is necessarily also similar at 10 ms and 15 ms, background no…