Harman preference curve for headphones - am I the only one that doesn't like this curve?

[thewas]

David Griesinger describes a nice method how to equalise headphones for frontal imaging similar to nearfield monitoring taking into consideration the individual HRTFs.

He has also recently announced a more user friendly software to achieve that
http://www.davidgriesinger.com/

 

[bobbooo]

David Griesinger describes a nice method how to equalise headphones for frontal imaging similar to nearfield monitoring taking into consideration the individual HRTFs.

He has also recently announced a more user friendly software to achieve that
http://www.davidgriesinger.com/

This method produces incorrect results due to the SLD effect, as @Mad_Economist points out at the end of this post.

 

[thewas]

This method produces incorrect results due to the SLD effect, as @Mad_Economist points out at the end of this post.

In the end any approach trying to imitate loudspeakers in a room via headphones is strongly limited and compromised. At least this one gives a nice frontal localisation for people who don't like the usual median one.

 

[dasdoing]

The definition of a flat speaker (in terms of translating it into a headphone experience) implies that you'd be listening to the speaker in a room rather than an anechoic chamber, or do you do most of your speaker listening in an anechoic chamber!?

you were defining neutral, not a speaker in a room. a perfectly flat speaker is not neutral in a room.
actualy we are not even at the point were we can clearly meassure neutrality of a speaker in a room. the psychoacoustics of this is not clear yet. there are papers showing that we can't meassure above the 1000Hz-ish region with a 500ms window, like we do in the bass region.
but one thing we can meassure and get to neutrality: an extrem nearfield setup (we exclude the reflections).
now, simulating a room on headphones with FR is not possible at all. you need binaural systems for that.
neutrality has clearly only one definition: the FR of the recorded material is the same as the reproduced one

This method produces incorrect results due to the SLD effect, as @Mad_Economist points out at the end of this post.

I don't understand the point of SLD in this context. they talk about SLD in the context of creating a standard. we are talking about individual response here

 

[dasdoing]

He has also recently announced a more user friendly software to achieve that
http://www.davidgriesinger.com/

I wonder if this is good for checking neutrality of in room house curves, too

 

[Robbo99999]

you were defining neutral, not a speaker in a room. a perfectly flat speaker is not neutral in a room.
actualy we are not even at the point were we can clearly meassure neutrality of a speaker in a room. the psychoacoustics of this is not clear yet. there are papers showing that we can't meassure above the 1000Hz-ish region with a 500ms window, like we do in the bass region.
but one thing we can meassure and get to neutrality: an extrem nearfield setup (we exclude the reflections).
now, simulating a room on headphones with FR is not possible at all. you need binaural systems for that.
neutrality has clearly only one definition: the FR of the recorded material is the same as the reproduced one



I don't understand the point of SLD in this context. they talk about SLD in the context of creating a standard. we are talking about individual response here

Well, whatever, I'm fed up with arguing semantics, etc, do what you like to your headphones, not my business, I don't have to agree with it.

 

[bobbooo]

I don't understand the point of SLD in this context. they talk about SLD in the context of creating a standard. we are talking about individual response here

The SLD effect does apply to individuals, and must be taken into account when producing any kind of headphone compensation (whether individualized or standardized), if not the results will be incorrect. From Theile's paper:

Our own experiments and similar work carried out by Mathers and Landsdowne [15] show that differences are found in the sound pressure levels measured in the auditory canal in the range of 7.5-14 dB between loudspeaker and headphone reproduction, while the test subject considers both to have equal loudness. Headphone reproduction requires more sound level in the auditory canal than equal loudspeaker reproduction.

 

[Maiky76]

To the extent you have told us Oratory creates his EQ using his ears, then cognitive bias and other nuisance variables like him being in headphone business (and liking what they design) will loom large with him. So if you want controlled results, you don't have it in him.

As for "smoothing" the results, here are his AKG K371 filters:

View attachment 109247

Are you kidding me? Q of 0.71? Not 0.7, but 0.71??? You really think he could reliably tell the difference in a controlled test between Q of 0.71 on that high shelf and 0.7? Research shows our sensitivity there to be a few dBs, not one hundredth of a dB.

All of his filters have too much precision. Look at all the ones with gain of -1.9. Or 1.8. These are all like someone sticking a wet thumb in the air and saying the temperature is 33.1 degree!

These look mechanically created to me. I can't imagine him sitting there and futzing with 0.1 Q variations and arriving at anything valid. The amount of time it would take to develop such would be huge anyway.

Look at my EQ for example of perceptually created filters using measurements as guide:

See how there are no fractions on the Q factors? And how few filters there are?

I can guarantee you that my 3 and 4 dB corrections above are audible to all sighted or blind. The 1.5 dB one is tougher and you could leave it out if you like. There is no way you can say that about his filters as I explained above.

Are we to believe that with all their research, Harman created the K371 to need 9 filters to match their preference target as Oratory created? Doesn't pass the smell test, does it?

BTW, there are also side-effects from these filters in that not all implementations do what you see in the pretty graphs. Filters can have overshoot before and after that is not visible in the UI. This is another danger in auto-generating filters.

Bottom line is what I said: he is over filtering because he is chasing a mechanical target with measurements that are too variable to justify such. Averaging a few headphones doesn't help in that regard because each measurement is variable in itself. What helps is understanding the nature of variability and making proper judgement in developing filters as I try to do (using psychoacoustics and strict listening protocols).

I have a few comments:
@Sean Olive
- If I am not mistaken, Harman used "Virtual Headphones" in one of their experiments to remove some of the biases i.e. touch information, comfort, weight etc. that could easily be detected by the user when blind but in direct contact with the DUT.
How many EQ points were used? Harman might have been using a FIR (which is even more complex), to simulate a number of headphones through a single one.
How close to the actual headphone *curve* or "representation" (whatever that may be through averaging smoothing etc.) of the performance of the real device did it need to be for the virtual headphone to be statistically indistinguishable from the real thing?
This should give a good indicator for the EQ complexity, parameters and tolerance we should be aiming for and probably a way to get a better representation of the performance of the DUT.

- If the Harman curve it to be trusted and we can characterize the performance of the device adequately,
there is no point EQing to anything else for a general EQ. I might restate what is already known there...
No matter what efforts you put in designing your EQ it will have your own subjectivity (taste) signature built in.
You are your own data point, it is the most important for you, I do realize that, but this is it, no more, no less.
Even trained listeners must be averaged.
Your EQ may work for some but it may not for others then we are back arguing on what sounds "good" to each subjectivity i.e. taste.
It is interesting in itself but not that fruitful, "y" prefer Bordeaux wine "x" prefers Burgundy wine Ok, then what?
Why should be argue that one is better than the other? Give AI a few years, and maybe we may see something coming from it...

The default Harman curve on the other hand is speaking for a much larger portion of the population (circa 64%) so statically if the population is large and representative enough we have a far greater chance that it will provide the user with something that hits the spot.
Nothing prevents the user from adding its subjective preference which will always be the case anyway but most of the leg work is done upfront.
I would argue that if someone is interested in using any of the published EQs that person will most probably deviate form the default settings and invariably add that person's subjectivity into the mix; if it is not just to compensate for perceived unit to unit variations.

- With regards to the "0.71 vs 0.7" comment I am not arguing that it may or may not be audible in fine however I believe this has more to do with the representation of the filter than its audibility. To me this is a different problem, not related to any EQ or designer in particular.
This is probably pedantic but let's take the following example:
BQ1 PEQ Freq = 1000.0, Gain = 3.00, Q = 3.00,...
BQ2 PEQ Freq = 1000.0, Gain = 4.00, Q = 1.00,...

If I were to express the biquads in their direct form :
bb1 =
1.003749261367279 -1.977580181386317 0.978074156083333
aa1=
1.000000000000000 -1.977580181386317 0.981823417450612
bb2 =
1.014808395051763 -1.945190025905157 0.934555368306628
aa2 =
1.000000000000000 -1.945190025905157 0.949363763358392

Your argument would be: why do you need so many digits, you can't hear the difference between
-1.945190025905157 0.934555368306628 and
-1.977580181386317 0.981823417450612, well you might.

This is equivalent to the spelling for written languages, the more correct it is the more precisely the meaning is conveyed, nothing else.
You are right when using computer generated EQ we tend to over do things, I try not to (and probably fail) but you need precision to convey the design.

 

[Robbo99999]

I have a few comments:
@Sean Olive
- If I am not mistaken, Harman used "Virtual Headphones" in one of their experiments to remove some of the biases i.e. touch information, comfort, weight etc. that could easily be detected by the user when blind but in direct contact with the DUT.
How many EQ points were used? Harman might have been using a FIR (which is even more complex), to simulate a number of headphones through a single one.
How close to the actual headphone *curve* or "representation" (whatever that may be through averaging smoothing etc.) of the performance of the real device did it need to be for the virtual headphone to be statistically indistinguishable from the real thing?
This should give a good indicator for the EQ complexity, parameters and tolerance we should be aiming for and probably a way to get a better representation of the performance of the DUT.

- If the Harman curve it to be trusted and we can characterize the performance of the device adequately,
there is no point EQing to anything else for a general EQ. I might restate what is already known there...
No matter what efforts you put in designing your EQ it will have your own subjectivity (taste) signature built in.
You are your own data point, it is the most important for you, I do realize that, but this is it, no more, no less.
Even trained listeners must be averaged.
Your EQ may work for some but it may not for others then we are back arguing on what sounds "good" to each subjectivity i.e. taste.
It is interesting in itself but not that fruitful, "y" prefer Bordeaux wine "x" prefers Burgundy wine Ok, then what?
Why should be argue that one is better than the other? Give AI a few years, and maybe we may see something coming from it...

The default Harman curve on the other hand is speaking for a much larger portion of the population (circa 64%) so statically if the population is large and representative enough we have a far greater chance that it will provide the user with something that hits the spot.
Nothing prevents the user from adding its subjective preference which will always be the case anyway but most of the leg work is done upfront.
I would argue that if someone is interested in using any of the published EQs that person will most probably deviate form the default settings and invariably add that person's subjectivity into the mix; if it is not just to compensate for perceived unit to unit variations.

- With regards to the "0.71 vs 0.7" comment I am not arguing that it may or may not be audible in fine however I believe this has more to do with the representation of the filter than its audibility. To me this is a different problem, not related to any EQ or designer in particular.
This is probably pedantic but let's take the following example:
BQ1 PEQ Freq = 1000.0, Gain = 3.00, Q = 3.00,...
BQ2 PEQ Freq = 1000.0, Gain = 4.00, Q = 1.00,...

If I were to express the biquads in their direct form :
bb1 =
1.003749261367279 -1.977580181386317 0.978074156083333
aa1=
1.000000000000000 -1.977580181386317 0.981823417450612
bb2 =
1.014808395051763 -1.945190025905157 0.934555368306628
aa2 =
1.000000000000000 -1.945190025905157 0.949363763358392

Your argument would be: why do you need so many digits, you can't hear the difference between
-1.945190025905157 0.934555368306628 and
-1.977580181386317 0.981823417450612, well you might.

This is equivalent to the spelling for written languages, the more correct it is the more precisely the meaning is conveyed, nothing else.
You are right when using computer generated EQ we tend to over do things, I try not to (and probably fail) but you need precision to convey the design.

I like those questions of yours to Sean Olive, useful in finetuning EQ approach.

 

[dasdoing]

The SLD effect does apply to individuals, and must be taken into account when producing any kind of headphone compensation (whether individualized or standardized), if not the results will be incorrect. From Theile's paper:

Sorry, I might be to stupid to understand, but I read this before and don't see the problem. If "Headphone reproduction requires more sound level in the auditory canal than equal loudspeaker reproduction" then the mentioned test will result in more sound level in the auditory canal than equal loudspeaker reproduction

 

[Dreyfus]

Theile states that the reproduction of sound through headphones is flawed since the possibility for adaption and orientation in front of a real sound source with the natural HRTF excitation plus the psychoacoustic filtering is missing.
His reasoning to prove that the loudness equalization must be wrong is based on the fact that the loudness transfer function and the physical transfer function measured right in front of the ear drum do not match.

I am not sure if the latter is a valid argument. When we find that listening to loudspeakers causes different acoustic and psychoacoustic phenomena compared to listening to headphones, why would you still expect the absolute level at the ear drum to be the same? Even when we recreate the exact transfer function at the listeners ear drum - solely based on absolute measurements - the shortcomings of headphone listening (lack of spatial orientation, physical bass sensation etc.) would still be there. If we look at it that way, headphone listening is fundamentally wrong per se and cannot be fixed by a simple equalization model.

As already mentioned in other threads, I have had good success with loudness tuning after applying generic headphone equalization presets based on GRAS gear measurements so far. After all, a microphone can measure but not evaluate sound.

Just my 2 cents.

we're not supposed to hear all frequencies to the same perceived level

Still, the community expects headphones to perfectly match the Harman target which does only render the broad distribution of energy but not HRTF-related high Q notches that our hearing expects for natural sound sources. If we look at it that way, a perfect Harman match is just as wrong as a loudness equalization that tries to achieve a continuous auditory threshold across the whole spectrum (in this specific case: treble response).

 

[Robbo99999]

Still, the community expects headphones to perfectly match the Harman target which does only render the broad distribution of energy but not HRTF-related high Q notches that our hearing expects for natural sound sources. If we look at it that way, a perfect Harman match is just as wrong as a loudness equalization that tries to achieve a continuous auditory threshold across the whole spectrum (in this specific case: treble response).

Try the different approaches and see which fits best. I'm happy with EQ's based on GRAS measurements done to the Harman Curve, that fits me very well. If that relatively easy approach doesn't suit you then it's certainly worth trying the other methods......there is no perfect solution currently.....but some solutions are better than others for different people, like if you happen to differ in ear & head physiology a long way off from the average on which the Harman Curve is based, then fine you won't like it, then it's worth trying other approaches. I may try the Impulcifier approach sometime (https://github.com/jaakkopasanen/Impulcifer), but I don't have a great deal of driving force for me to do it due to me already liking the Harman Curve, but this Impulcifier approach should give me better spatial representation in headphones....I might get round to it sometime, but I'll be sending my AKG K702 off to Oratory soon so he can measure them......perhaps with a Harman EQ based exactly on my specific unit of headphone then this will drive the Impulcifier idea even further into the background.

 

[dasdoing]

headphone listening is fundamentally wrong per se and cannot be fixed by a simple equalization model.

there is no doubt it is. it is just not how we hear real sounds

 

[m8o]

With all these headphones I really hate the Harman curve (please don't delete my account @amirm )
.
.
.
Am I alone here in really not liking Harman preference curve? Like really not liking it

Wish I saw this when you posted this so I could be the 1st reply'er to say... I'm 100% with you. My foray into AutoEQ yesterday only managed to reinforce this same opinion I had formed years past as the Harman target evolved, rather than giving me an epiphany to the virtues of it or that I had that wrong.

Just like anything created by empirical committee, a little under 1/3 of the population are outliers at or beyond 1-standard deviation from the norm of the bell shaped curve; and around 5% of the population are two standard deviations outside the norm; that's us! lol. And if -we- were in the committee of people that helped form the Harman curve, -we- would have been the outliers from the preferrence / opinion that formed the norm / target.

That said, it is only certain aspect of it I do not like (safe to say, I detest). That being the extend of the midrange peak and the deep deep rolloff of the highest highs. I am going to give AutoEQ's --sound_signature feature a try to see if I can successfully a) utilize target curves, b) but altered to my preference.

 

[devopsprodude]

Problem is that you aren’t actually using the Harman curve. You’re using someone’s implementation of it and there are sometimes up to like ten different implementations to choose from for a particular headphone model. When using SoundSource on Mac, I got significantly different results depending on which K712 profile I chose, and ended up liking the one labeled "Reference Audio Analyzer" although Crinacle's profile was also good to my ears.

For me the profile is a good starting point, and then I overlay 31 band eq over that to get things how I like (I tend to prefer elevated bass AND treble, which according to Sean Olive is often associated with people who play video games).

 

[Sean Olive]

Wish I saw this when you posted this so I could be the 1st reply'er to say... I'm 100% with you. My foray into AutoEQ yesterday only managed to reinforce this same opinion I had formed years past as the Harman target evolved, rather than giving me an epiphany to the virtues of it or that I had that wrong.

Just like anything created by empirical committee, a little under 1/3 of the population are outliers at or beyond 1-standard deviation from the norm of the bell shaped curve; and around 5% of the population are two standard deviations outside the norm; that's us! lol. And if -we- were in the committee of people that helped form the Harman curve, -we- would have been the outliers from the preferrence / opinion that formed the norm / target.

That said, it is only certain aspect of it I do not like (safe to say, I detest). That being the extend of the midrange peak and the deep deep rolloff of the highest highs. I am going to give AutoEQ's --sound_signature feature a try to see if I can successfully a) utilize target curves, b) but altered to my preference.

.


The Harman Target Curve is only a guideline that seems to satisfy the majority of tastes as you pointed out..It seems to satisfy both trained and untrained listeners.

Above 10kHz we don't even include headphone measurement data in calculating predicted preference scores as there are too many uncertainties in couplers, repeatability in positioning of the headphones on pinnae and humans, and headphone-pinnae interactions which can vary significantly between individual pinnae and headphone design.. To really nail down preferred response above 10 kHz would probably require purchasing the GRAS high resolution coupler/microphone and probably screening of listeners using HF audiometry because I suspect the preferred response is highly dependent on hearing loss/sensitivity and other factors..

 

[tili]

I just got my ears vacuumed, had a lot of wax buildup and one ear was getting blocked. Now Harman Target Curve feels cold
I used to like it. Just gotta adjust the bass to my preference probably.
What was the low shelf frequency they were using when they conducted the preference test with differing bass preferences?

 

[markanini]

That's because your brain needs time to adapt to the change in hearing.

 

[solderdude]

yes, took me a day to get used to the 'sharp' sounding world after my ears were flushed the first time.
After a while everything sounded 'normal' again.
It is amazing how the brain adapts to its 'sensors'

 

[solderdude]