I'm a bit out of my depth here but the "after" plot looks a bit extreme to me, what about transient response?  Are we to believe that this method makes the driver so fast you can't even detect it anywhere?  The "after" plot here looks to me like there is no sound being produced at all.  Even the cleanest, flattest headphone is going to have a plateau at the very beginning right? 

I could be missing something, I have no idea what minimum phase FIR is.

I'm pretty sure there was no sound being produced/recorded from the headphones in the second graph. That said ringing/resonance does go hand in hand with FR peaks/dips.

This shows that 90% of the "ringing" to be simply a direct result of the phones' FR imperfections...

I'm pretty sure that's not how it works

That filter extends into the time domain and it's not just changing the FR.  It's adding extra content to the signal in order to cancel out things the driver will do, kind of like active noise cancellation.

AFIK that kind of thing can work very well when you just use a short test signal and then convolve the test signal to exactly counter the problems.  It doesn't work out when you try to derive a generally applicable FIR filter that will work with any input signal because the drivers aren't quite linear enough to match the mathematical models and being just a little bit off will make things sound worse.

To make that idea work you'd have to do something silly like record your headphones playing each individual track in order to get the FIR filter to fix it.  Only slightly more practical would be modding your headphones with a "feed forward" mic like the good ANC cans have and running it's output back to you PC where you could run some kind of DSP engine that could do all the math in real time.

Also, can you link to that post?  I'm not even sure if that second one was actually recorded or if the data from the first graph was just processed and then redisplayed.

I agree that it is more practical in the long run to mod the headphones. But I don't think that was an echo canceller or predictor structure. That more than likely was straight FIR convolution to the impulse response which did not took into consideration non-linear distortion and positional variations.

Oops.  I didn't mean to give the impression that I thought xnor did it that way.  It seemed pretty clear to me that it was a straight FIR convolution.

I was just throwing out an idea that worked on a similar principle.

It's is very interesting though. I should apply corrective EQ with various filters and take measurements.

You should start with an Ultrasone or the HD 700.