[xnor]

Why are you interested in phase at 20 kHz? I don't think there's even a point going above 10 kHz. The lower kHz range might be interesting if there are crossovers and bass-mids for reasons mentioned above.
 
On the other hand, it might be another thing some people will point at and say: uww there's some phase shift, must sound bad! So maybe we're better off without that extra data.

 

[jaddie]

Quote:

Why are you interested in phase at 20 kHz? I don't think there's even a point going above 10 kHz. The lower kHz range might be interesting if there are crossovers and bass-mids for reasons mentioned above.
 
On the other hand, it might be another thing some people will point at and say: uww there's some phase shift, must sound bad! So maybe we're better off without that extra data.

I'm not really interested in 20KHz phase, and I agree, there isn't any point at the top end, but it shows there are inaccuracies built in which while progressively smaller with lower frequencies are still around.  It makes for challenges with repeatability. If I were publishing test results I'd have a problem showing data with that potential for error, and would probably pick my battles. 
 
Your second point...yup, no kidding.  Since a phase plot is yet another example of a hypersensitive representation of inaudible issues, it's probably best left off.  

 

[ultrabike]

Quote:

Yeah, that's the problem: to do it real time you'd have to know the distance between the driver and ear to start. There are other ways (through calculations on the impulse response) but I'm not able to do that at the moment.

 
Not perfect but...
 
Impulse Response (AKG K701):

Frequency Response Magnitude from Impulse Response (uncompensated)

Frequency Response Phase from Impulse Response (uncompensated)

 
Sampling rate seems to be about ~170kHz for the IR.
 

 
The slope is a function of the number of "zeros" to the left of the main peak (as expected):
 
IR (less "zeros" to the left of main peak - attempting to artificially remove distance and delay):

FR Magnitude (after removing IR "zeros" to the left of main peak):

FR Phase (after removing IR "zeros" to the left of main peak):

Quote:

Why are you interested in phase at 20 kHz? I don't think there's even a point going above 10 kHz. The lower kHz range might be interesting if there are crossovers and bass-mids for reasons mentioned above.
 
On the other hand, it might be another thing some people will point at and say: uww there's some phase shift, must sound bad! So maybe we're better off without that extra data.

 
I tend to agree, but not 100% sure. Specially if there is some really weird issues and phase jumps bellow 10kHz... Dunno.

 

[stv014]

Another attempt, using this (not particularly good quality, but suitable for this example) impulse response:
   
 
First, I calculate the phase delay vs. frequency (note: the vertical scale is actually in ms, not dB):
   
On the right, the top end of the response is shown, zoomed in. Its purpose is to find the delay of 0.00056245 s that needs to be compensated for the phase response to go back to zero at the end.
 
Now, with that time offset, here is the group delay (top left), phase delay (top right), and phase (bottom left, this time 1 dB on the vertical scale is 10 degrees of phase):
   
 
   
At the bottom right, there is a second phase graph that is offset by one cycle (360 degrees), since that gives a more realistic response for a filter that is assumed to be close to minimum phase.

 

[ultrabike]

Interesting! What headphones are those? Could the notch at around 200Hz be the cause of the phase slope change at around that same frequency?
 
The AKG seems well behaved till around 2 kHz. The phase response slope change afterwards might be due to the FR emphasis later on. Since this is uncompensated, that slope change might be necessary to account for headphone + ear coupling...
 
Again... Dunno...

 

[stv014]

It is a DT770 Pro 250. The phase response does tend to depend on the magnitude response, in fact, for a minimum phase filter, it can be calculated from the magnitude response.
 
For comparison, here is the group delay and phase (the right channel is "buggy" and off by -360 degrees) graph for a version of the above impulse response that has been made minimum phase. There are some similarities, but also differences:
   
Also, yet another phase response for the original IR, this time it starts at +360 degrees, and ends at 0 degrees (by changing the assumed delay of the impulse to 0.00054162 s):

 

[ultrabike]

Yes, a simple delay in the signal corresponds to a different phase response (average) slope. Maybe one could assume and standardize a certain distance from start to the IR peak to be more consistent.
 
Quote:

The phase response does tend to depend on the magnitude response, in fact, for a minimum phase filter, it can be calculated from the magnitude response.

 
xnor reminded me of that at some point in the past, and it is now burned into my skull

 

[stv014]

Quote:

Yes, a simple delay in the signal corresponds to a different phase response (average) slope.

 
That is why I suggested measuring the group delay instead, it is not sensitive to overall delay (the whole graph just gets shifted up or down by the amount of delay, but that is a non-issue because it is the variation in delay that matters), and the unwrapped phase response being "off" by a whole cycle, like on some of the graphs above, is not a problem either. Also, it may be more intuitive to show how much a band of frequencies is delayed relative to others, and it does not make the high frequency response look worse than it really is.
However, especially at mid to high frequencies, a CSD would probably be the most useful.

 

[jaddie]

Quote:

 
That is why I suggested measuring the group delay instead, it is not sensitive to overall delay (the whole graph just gets shifted up or down by the amount of delay, but that is a non-issue because it is the variation in delay that matters), and the unwrapped phase response being "off" by a whole cycle, like on some of the graphs above, is not a problem either. Also, it may be more intuitive to show how much a band of frequencies is delayed relative to others, and it does not make the high frequency response look worse than it really is.
However, especially at mid to high frequencies, a CSD would probably be the most useful.

Perhaps group delay with some sort of normalized scaling?  But I like the idea.  And I'll vote for CSD also/instead of. 
 
edit: By Normalized, I mean something that minimized the amount of delay and highlights the rate-of-change.