Open Nuopel opened 7 years ago
1-2 The SNAC pitch algo is implemented however the results are quite similar to the modified cepstrum: the low pitches midi are accurately founded but the high are not wrong pitched. 3 warped function imply an accurate selection time of the notes thus for a later time
1 I make the hypothesis that the problem in high frequency is that the signal is not accurate enough to locate the peak as in high frequency the number of points by peaks decrease.
The first idea was to zero pad the signal but it's stupid as autocorrelation and the cepstrum are time domain based.
The second idea is to upsample the signal using first a linear interpolation on two , then may be a second order polynomial interpolation with 3 points. LETS DO IT IN PYTHON YEAHHH (i think an autocorrelation, using vector and wav should be fine for a beginning ).
Implement SNAC pitch algorithm.
Try a dual pitch algo : |low frequency: Modified Cepstrum, |high frequency: SNAC. ->filtering?
Test warped function to see if pitch results are better.
Find others onset detections.