Open andresbrocco opened 1 year ago
Indeed, this normalization behavior is redundant nor it is following the original paper that the algorithm cites (Tolonen & Karjalainen, 2000), and I cannot find a reason this normalization was added in the first place. A pull request with a fix is welcome.
PercivalBpmEstimator is the only Essentia algorithm using generalized auto-correlation, so we should double check that its behavior is not affected.
Also test_autocorrelation.py
is missing a test for the generalized case.
@dbogdanov I don't think this issue should be labeled "algorithms QA". Instead, it should be on "bug".
I'm currently studying the library, and I've found out that the implementation of the 'AutoCorrelation' output is not consistent between its options:
If you use 'generalized=false', it calculates the autocorrelation correctly. If you use 'generalized=true', the overall magnitude of the autocorrelation will be scaled by 1/sizeFFT^k, which is extremely bizarre (its scale should not depend on the size of the FFT).
Check out the current source code: