First I have to say that these drawing are gorgeous! but I know it is not the point
Second, @jacob.hart will know how his patch works better than me - I’m looking for it now but I don’t seem to know where to find it - we did build a huge database of code so can you help me
I am mostly curious in how pitch mean (at best 2d) is managing to be postprocessed. MFCCs will have to go through a dimension-reduction algo but maybe the said algo is freaking out on 2d-to-2d - or maybe it is expected behaviour and you should bypass that step somehow.
Looking forward to hearing from either of you so I can help a little more
in this patch, the left side is how i had it before and the right side is just bypassing the fluid.umap~not sure if this still needs a step before the fluid.normalize~ so that each point receives two values?
anyway, the hallucinations seem to be gone, but this was analysed with “pitch loudness mean mean” so this looks wrong too. third pic is what comes into the fluid.normalize~, not sure if the pointsize is correct there?
you seem to have only one dimension in your dataset. Check (print) the dataset in which you send the analysis. Also, ping @jacob.hart on his github, there might be a bug in his wrappers. or in your understanding of them, if you can’t request 2 things for instance.