sevagh/pitch-detection

Clean up the normalized FFT vs time-domain FFT

Closed this issue · 2 comments

sevagh commented

It's a bit messy -

  • The time-domain autocorrelation does the opinionated improvements of MPM
  • The FFT autocorrelation doesn't have the same but it has some form of normalization via completely pointless division by result[size/2]

Come up with:

enum NormalizationStrategy

Make sure they align for time domain and FFT.

sevagh commented

You can see autocorrelation gets 249Hz on the E2 with -DPORTABLE_XCORR, which are the advantages from NSDF: https://github.com/sevagh/mcleod#83hz-open-e-string-on-guitar-1

With FFTS, autocorrelation gets 282Hz on the E2, proof of non-normalization: https://github.com/sevagh/mcleod#83hz-open-e-string-on-guitar

sevagh commented

This is done - there's basically no NSDF here, which doesn't matter for MPM or YIN - NSDF makes autocorrelation a little less bad, but still not good. So, I've removed all implicit normalization.