• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Audibility of Equalizer Ringing

jaakkopasanen

Member
Joined
Jul 12, 2020
Messages
87
Likes
343
I conducted a remote AB listening test looking for audibility of equalizer frequency response peaks using abxtests.com. This was done to check if peaks in equalizer frequency response are harmful and how they should be regularised (see Reddit post for more info). The different dB values here correspond to the maximum allowed slope steepness (first derivative) of the equalizer frequency response meaning that higher values allow steeper slope and therefore higher peaks. The hypothesis was that allowing for steeper slopes makes it possible to equalize more accurately but at the same time the peak will start to resonate (ring) audibly and that people would prefer some kind of compromise between the two.

In total 10 subjects finished the test, each using different pair of headphones measured by oratory1990. Soon it started to come clear that deviation from target due to peak regularisation nor the filter resonance are not problems as a few test subjects refused to finish the test as they felt they cannot hear the difference. We can see this also in the results as there clearly is no preference for any of the regularisation values in the table. The "Shapiro-Wilk P" column is the probability of the sample coming from normal distribution and we can see that all of these probabilities are extremely low. My idea for this statistical test was that if there is a compromise between the equalization accuracy and filter resonance, the results would exhibit a normal distribution.

This isn't a massive amount of data and the remote listening testing methodology isn't necessarily up to the highest scientific rigour but still I'm fairly confident in concluding that for most headphones equalized to Harman target the filter resonance is not an issue. Anecdotally I did myself however experience a need for compromise but I took the test using Custom Art FIBAE 3 which has massive ~20 dB gap between 5 kHz and 7 kHz so the issues were audible, only if very slightly for most songs.

Limited Slope EQ Test Results.png
 
OP
jaakkopasanen

jaakkopasanen

Member
Joined
Jul 12, 2020
Messages
87
Likes
343
Nice work, what are the main numbers in the table though that add up to 30?
It's the sum of each row.

:facepalm:
Just avoid eqing the notch for that matter.
To hear if the dip is there or not, use sine sweep like https://www.szynalski.com/tone-generator/
Don't put blind test on everything that's not needed. There are more effective ways of doing things.
This work was done for improving automatic equalization of headphones in AutoEq. "Just avoid eqing the notch for that matter." is a good advice but mathematical formulation of what is a notch that should be avoided and what is not is surprisingly difficult.
 

JohnYang1997

Master Contributor
Technical Expert
Audio Company
Joined
Dec 28, 2018
Messages
7,175
Likes
18,292
Location
China
It's the sum of each row.


This work was done for improving automatic equalization of headphones in AutoEq. "Just avoid eqing the notch for that matter." is a good advice but mathematical formulation of what is a notch that should be avoided and what is not is surprisingly difficult.
Sorry for this but AutoEQ just plain doesn't work well.
Or for the matter of practicality, just limit Q to a fairly low number and it's all good to go.
 
OP
jaakkopasanen

jaakkopasanen

Member
Joined
Jul 12, 2020
Messages
87
Likes
343
That was clear to me, but not the invidual numbers which create this sum, some kind of preference or passed trials in the AB?
That would be: how many times each option was selected by the test subjects as the most preferred one.
 

thewas

Master Contributor
Forum Donor
Joined
Jan 15, 2020
Messages
6,747
Likes
16,182
That would be: how many times each option was selected by the test subjects as the most preferred one.
Thank you, there seems to be some trend for preference on less eq gain (which is also the experience of others) but like you say it needs more research of course.
 

JohnYang1997

Master Contributor
Technical Expert
Audio Company
Joined
Dec 28, 2018
Messages
7,175
Likes
18,292
Location
China
There shouldn't be any gain over 0dBFS. Plain and simple. I don't understand what this all is.
 

thewas

Master Contributor
Forum Donor
Joined
Jan 15, 2020
Messages
6,747
Likes
16,182
There shouldn't be any gain over 0dBFS. Plain and simple. I don't understand what this all is.
That is obvious and achieved with the appropriate negative pre-gain (and was hopefully done here too) but not the question of this research.
 

wwenze

Major Contributor
Joined
May 22, 2018
Messages
1,284
Likes
1,827
That is obvious and achieved with the appropriate negative pre-gain (and was hopefully done here too) but not the question of this research.

As in, peak boosts are bad, peak attenuations are good. Nothing to do with clipping.

And there's a simple reason for that really. With any EQ you do (Assuming IIR i.e. the "normal" EQ; we put FIR filters aside for now), any shift in amplitude must come with a corresponding shift in phase. It's just the unfortunate reality of mathematics. And ringing is one of the result of this phase alteration.

So if you want to screw the phase of the signal, you want to screw it at where it is inaudible, not where it is 10dB louder than the rest of the original signal.
 

AnalogSteph

Major Contributor
Joined
Nov 6, 2018
Messages
3,334
Likes
3,278
Location
.de
FIR actually is much worse in that regard than IIR, must be the pre-ringing. If you want to try it out, I think Foobar2000 still ships with the same standard EQ plugin based on Shibatch Super Equalizer to this day... the Winamp version at least was creating nice steep steps between EQ bands in the frequency response (I tried it with the RMAA test signal at one point), and we all know what happens in time domain if FR gets all kinky like that. More than 2 dB worth of step started making things sound noticeably worse, I thought at the time.

In case of IIR EQ, if what you are compensating for is a minimum phase phenomenon, you'll even out the phase along with the frequency response. Any ringing you are introducing is just canceling some of the exact opposite polarity in the original response. This will obviously not work properly for any linear phase issues that are purely rooted in time delay... as the odd dip turns out to be. So basically you really need both approaches at the same time, but IIR quite arguably has higher priority as compensating for dips is fraught with peril after all.
 

thewas

Master Contributor
Forum Donor
Joined
Jan 15, 2020
Messages
6,747
Likes
16,182
As in, peak boosts are bad, peak attenuations are good. Nothing to do with clipping.

And there's a simple reason for that really. With any EQ you do (Assuming IIR i.e. the "normal" EQ; we put FIR filters aside for now), any shift in amplitude must come with a corresponding shift in phase. It's just the unfortunate reality of mathematics. And ringing is one of the result of this phase alteration.

So if you want to screw the phase of the signal, you want to screw it at where it is inaudible, not where it is 10dB louder than the rest of the original signal.
I know that postive minimal phase PEQ means adding oscillators and thus ringing in time domain which is one of the reasons they should be avoided if possible or used with care (low gain and Q factor), but this was also the purpose of this test to test their audiblity.
 

fluufy

Member
Joined
May 7, 2021
Messages
36
Likes
49
I did my own totally unscientific experiment - I hacked autoeq to increase the allowable slope to 24dB/octave. This was while making a custom eq for B&O Beoplay H9 3rd gen, using Bluetooth with ANC on (the main use-case for a headphone like this).

There are no good measurements for these, and my own effort at measuring by bodging a flat-plate with a umik-1 was even worse, so I did it by ear using sine sweeps. The issue was the slope between a 7dB dip at 3k, and a 4dB peak at 4.5k. If I allowed autoeq to chop off the peak they sounded flat, and the dip was necessary to stop female vocals from screeching.

There may well be ringing, but not that I can hear over BT with ANC on. In fact, using wavelet app these now sound better than I have been able to achieve with my husband’s Sony XM3’s. To my well-used ears anyway, far too many warehouse parties in the 90’s… Chart output from autoeq below:
38976F09-3310-4E73-B900-C26FCB5620BA.png
 

daftcombo

Major Contributor
Forum Donor
Joined
Feb 5, 2019
Messages
3,687
Likes
4,068
I conducted a remote AB listening test looking for audibility of equalizer frequency response peaks using abxtests.com. This was done to check if peaks in equalizer frequency response are harmful and how they should be regularised (see Reddit post for more info). The different dB values here correspond to the maximum allowed slope steepness (first derivative) of the equalizer frequency response meaning that higher values allow steeper slope and therefore higher peaks. The hypothesis was that allowing for steeper slopes makes it possible to equalize more accurately but at the same time the peak will start to resonate (ring) audibly and that people would prefer some kind of compromise between the two.

In total 10 subjects finished the test, each using different pair of headphones measured by oratory1990. Soon it started to come clear that deviation from target due to peak regularisation nor the filter resonance are not problems as a few test subjects refused to finish the test as they felt they cannot hear the difference. We can see this also in the results as there clearly is no preference for any of the regularisation values in the table. The "Shapiro-Wilk P" column is the probability of the sample coming from normal distribution and we can see that all of these probabilities are extremely low. My idea for this statistical test was that if there is a compromise between the equalization accuracy and filter resonance, the results would exhibit a normal distribution.

This isn't a massive amount of data and the remote listening testing methodology isn't necessarily up to the highest scientific rigour but still I'm fairly confident in concluding that for most headphones equalized to Harman target the filter resonance is not an issue. Anecdotally I did myself however experience a need for compromise but I took the test using Custom Art FIBAE 3 which has massive ~20 dB gap between 5 kHz and 7 kHz so the issues were audible, only if very slightly for most songs.

View attachment 151392
Looks interesting, but can you tell us how to read the chart, and conclusions?
 

bennetng

Major Contributor
Joined
Nov 15, 2017
Messages
1,634
Likes
1,692
I think Foobar2000 still ships with the same standard EQ plugin based on Shibatch Super Equalizer to this day... the Winamp version at least was creating nice steep steps between EQ bands in the frequency response (I tried it with the RMAA test signal at one point), and we all know what happens in time domain if FR gets all kinky like that. More than 2 dB worth of step started making things sound noticeably worse, I thought at the time.
I wondered why very few people mentioned this. My dead geocities website showed exactly the same thing:
 
Top Bottom