Real-Time Polyphonic Pitch Detection on Acoustic Musical Signals
Tom Goodman, Ian Batten · 2018
This paper presents an algorithm for fundamental frequency detection on polyphonic acoustic musical signals, based on a new `raking' method over the frequency-domain spectra. The algorithm is evaluated as a classifier, and boasts a good accuracy (83.20%) compared to other such methods, as well as the ability to function effectively in real-time, with a running-speed below 140ms per window evaluated. This proves to be real-time for the use-case, as the latency between an auditory stimulus and its perception by a person has been shown to be longer than this. The algorithm itself runs in linear-time, but is thus slowed by the O(nlog(n)) Fast Fourier Transform during preprocessing. Though the algorithm fails to account for certain edge-cases with overlapping harmonics as well as certain instruments, future work and improvements are also presented, paving the way for further research.