View Single Post
Old 01-19-2018, 09:12 AM   #30
geraintluff
Human being with feelings
 
geraintluff's Avatar
 
Join Date: Nov 2009
Location: mostly inside my own head
Posts: 346
Default

Quote:
Originally Posted by jcjr View Post
If stacking and overlapping numerous bands one would be cautious of the cumulative phase response maybe getting so extreme that it somehow affects the quality of the music. But sometimes the phase plot can get very weird indeed without (to my ear) adversely affecting the music.
An interesting property of minimum-phase filtering (which is what everything uses, except for "linear" EQs like FIR, or explicit phase-delay filters) is that the phase response is completely linked to the amplitude.

That is: it doesn't matter if you stack filters, overlap them, etc. - if the resulting amplitude response is the same, the phase is also the same. This is why if you add a boost/cut pair that cancel each other out in ReaEQ, the phase is also flat. In a very hand-wavey sense, strong phase-shifts happen when you have sharp cutoffs in amplitude.

There's even a way to take an amplitude response, and calculate the phases from it -
at which point you can even generate the minimum-phase impulse response (FIR kernel). I implemented this a few months ago for a (now abandoned) effect - I'm happy to dig this up if it's useful for anything. Now I think about it, perhaps I could use this to export a minimum-phase "correction kernel" from Spectrum Matcher...

(This minimum-phase-kernel is one method for constructing reverb impulses from sine sweeps - you assume that real impulse response is minimum-phase, which is somewhat reasonable, and reconstruct it from the magnitude response alone. The long tail corresponds to extremely fine-grained details in the amplitude response which are linked to extreme shifts in phase.)

Quote:
Originally Posted by mschnell View Post
"Nonlinar approximation" (curve fitting) is a rather complex, but well known numerical algorithm (i.e. software program). (I once crafted such a software from scratch.)
Learning EQs like this is a hell of a problem, yeah. Finding a solution is definitely possible, but the challenge is getting it consistent in the face of small changes.

For my Spectrum Matcher, you give it a "model" to learn from, and then play it an input, and it compares the spectra. (The model doesn't have to be white or pink noise - I've mostly used it to timbre-match vocal takes recorded months and miles apart. Veto explained how to use it better than I did.)

It can also try to correct the difference using EQ. I originally had wanted to generate this EQ using an optimisation method, a bit like what you describe (mine was just naive hill-walking).

I still would like to do this at some point, but the lazier option was essentially to use 30 shelf filters (or however many bands you specify for the analysis), which is slower than a well-chosen 6-band EQ would be.

Last edited by geraintluff; 01-19-2018 at 09:25 AM.
geraintluff is offline   Reply With Quote