Open johnmgithub opened 1 year ago
Very nice comment. At what signal lengths did you experience artifacts?
I noticed them at widths below a few k samples, but I apply some hill shading on my spectrograms so that makes them more visible than they would be on a flat spectrogram. I'm using a minimum length for the mother wavelet of 64k now, that eliminates them completely. I mainly use the spectrograms for looking at impulse responses, so I'm applying scale normalisation with those which might also make the artefacts more visible.
The integer truncation of the index into the mother wavelet causes some artefacts for short time spans. Those can be reduced without performance penalty by generating the mother data at a multiple of size for sizes below some threshold and passing that multiple to the daughter wavelet multiplication to scale the step.