This is nice, and elegant. But when I think of the mfcc, isn’t the negative “extrema” just going down to the “baseline”? Or does it actually mean “more energy” but in the opposite phase or something?
My idea of making the abstraction was to create a “rolling window” so that any time you hit the button, you get the maxpool of the last n steps. The only thing is then is that you don’t get the “profile” of the current moment + n-1 steps forward, but rather the current moment and n-1 steps before it, right? It might be a better design to add another inlet where you can bang for a maxpool of the upcoming n steps with just a single frame output. That could be more useful in “sampling mode”.
Added sampling mode to the abstraction and updated Owen’s patch. I get basically the same results, though I trained on a much smaller dataset (so it’s probably better!).
I don’t know, but @groma knows a lot about this, and so does @weefuzzy so I reckon they’ll have a nice explanation for us.
One thing for sure: as you will normalise, as long as you keep the full range of the shape, aka your way of offsetting or my way of keeping the polarity, it won’t change anything. Each dimension is normalised on its range.