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

[Sean Olive]

A collaboration between Drop and HIFIMAN.

Looking at the measurements it looks like that headphone needs a large bass boost. That could be an issue it it causes a lot of distortion.

 

[Sean Olive]

Many thanks for the reply @Sean Olive. I realise that measuring an in-ear DF response would require a reverberation chamber. My question referred to these comments from your 2015 paper (p. 3, where you outline the conditions under which you conducted the listening experiments (I hope you don't mind my reproucing your words here):



I took this to mean that essentially the following process had been followed:

1. Steady-state response of Revel F208 in Harman listening room adjusted to be flat at the listening position (which presumably meant that the anechoic response of the speaker was not flat, but rather recessed in the bass/boosted in the treble.
2. In-ear measurement taken using HATS system.
3. In-ear measurement of test headphones taken on same HATS system.
4. EQ used to adjust the headphones' response so that it matched that of Step 1 when measured on the same GRAS system, above.

And that this headphone response had then provided the starting condition from which subjects were asked to adjust bass/treble shelving filters to taste.

Did I understand that correctly? If so, my question was simply why this particular starting condition was chosen. Was this supposed to be an approximationg of a flat DF response? Or was there a particular reason why a flat steady-state response of a loudspeaker in a listening room was chosen? And, more broadly speaking, why was a flat steady-state response (as opposed to any other particular starting condition) chosen?

I realise the study was a few years ago, so no expectation here for a detailed answer It's just a question that I'd wondered about since reading the study.



That's very interesting, thanks. It seems the aparent discrepancy between IEM and over-ear headphone bass preferences remains a bit of a riddle.

Close:
1. The Revel F208 is flat anechoically. In a listening room it has this smooth downward tilt. The adjustments of the bass and treble allowed the listener to remove the flattening and get back to the original unmodified target or modify it to taste.

The bass and treble levels were randomly set in each trial, and the listeners' task was to adjust them to taste. Don't get hung up on why it was flattened. Flat was simply one of the settings possible in the listeners' choice of adjustment. It turned out if the speaker is anechoically flat to begin with, little or no adjustment above 300 Hz is needed. That is good news because you can probably equalize your subwoofers and call it a day.

 

[Sean Olive]

Many thanks for the reply @Sean Olive. I realise that measuring an in-ear DF response would require a reverberation chamber. My question referred to these comments from your 2015 paper (p. 3, where you outline the conditions under which you conducted the listening experiments (I hope you don't mind my reproucing your words here):



I took this to mean that essentially the following process had been followed:

1. Steady-state response of Revel F208 in Harman listening room adjusted to be flat at the listening position (which presumably meant that the anechoic response of the speaker was not flat, but rather recessed in the bass/boosted in the treble.
2. In-ear measurement taken using HATS system.
3. In-ear measurement of test headphones taken on same HATS system.
4. EQ used to adjust the headphones' response so that it matched that of Step 1 when measured on the same GRAS system, above.

And that this headphone response had then provided the starting condition from which subjects were asked to adjust bass/treble shelving filters to taste.

Did I understand that correctly? If so, my question was simply why this particular starting condition was chosen. Was this supposed to be an approximationg of a flat DF response? Or was there a particular reason why a flat steady-state response of a loudspeaker in a listening room was chosen? And, more broadly speaking, why was a flat steady-state response (as opposed to any other particular starting condition) chosen?

I realise the study was a few years ago, so no expectation here for a detailed answer It's just a question that I'd wondered about since reading the study.



That's very interesting, thanks. It seems the aparent discrepancy between IEM and over-ear headphone bass preferences remains a bit of a riddle.

It is. Jude at Head-Fi has a theory that the extra 4 dB of bass can be explained by measurement artifacts related to these acoustic impedance interaction effects with IEC couplers and IE headphones. If you measure the IE headphones on the B&K 5128 he says they tend to have 3-4 dB less bass closer to the AE/OE target. He has moved to the B&K 5128 because the ear simulator is based on a large study of human subjects and more accurate in terms of acoustic impedance. That may be true but the GRAS people believe the evidence he showed of a Westone IE headphone measured on different IEC couplers was not well controlled. They were measured by different people with different insertion depths, some with ear canal extenders, some with pinna -- all uncontrolled variables that could explain the observed differences. Also the frequency used to normalize the curve was 1 kHz, too high because the couplers have resonators in them that operate down to 800 Hz. The IEC standard recommends normalization at 500 Hz or lower.

We have a B&K 5128 and measured a lot of headphones on all the test rigs. Not had time to analyze the data because of other research priorities.

 

[bobbooo]

Looking at the measurements it looks like that headphone needs a large bass boost. That could be an issue it it causes a lot of distortion.

Those measurements on the Drop site were by Jude of Head-Fi. Speaking to Oratory about them he agreed with me that he likely didn't get a good seal when measuring the headphones. Oratory's measurements show better bass extension. As for distortion, it's pretty low across the entire frequency range, not only according to Head-Fi's (90dB@1kHz) but also Innerfidelity's measurements. Subjectively, I don't hear any distortion at all boosting mine up to the Harman target - in fact it's some of the cleanest bass I've heard.

One of "huge" assumptions here that Oratory's (or any EQ correction) sample of the headphone is similar in response to yours, so that his EQ will work for your headphone. How many samples did he measure? One-two? How many samples is Sonar Works EQ based on?

My experience with measuring multiple samples of any given model of headphone is that the tolerances in manufacturing are simply too wide to make that assumption. You may get better quality control for the premium models above $600- $1k but not necessarily in sub $300 headphones.

So unless you have access to good measurements and can use them to EQ your own headphone, all bets are off in regards to what you are listening to.

This is an important (and often overlooked) point. Some companies' / models' unit variation does seem better than others though, sometimes belying their price. For example, as @Robbo99999 mentioned, Oratory, who's said most of his posted measurements and EQ profiles are based on multiple units by now, found the HifiMan HE4XX's unit variation to be very low. Direct quote:

I measured 3 specimens by now, including yours and one that Olaf (from OluvsGadgets) brought to the lab.
I built an average from all of them and use this for the EQ PDFs. They are all pretty close together actually, so much so that it doesn't make much sense to create an EQ setting specifically for one specimen only. Unit variation was excellently low on the three specimens I tested. Quite surprised by this!

That's a $180 planar magnetic headphone (despite planars usually being more expensive than dynamic driver models - I actually got mine for just $140 on promotion, and they've been as low as $120), which has low enough variation between units that individual EQ settings wouldn't even make sense, presumably due to this unit variation being around the same or less than placement variation.

Talking of placement variation, this seems to be another often ill-considered factor when it comes to using headphone measurements for developing EQ profiles, sometimes only a single-placement (and single-unit) measurement being used to base EQ settings on for all users and their units, when in reality every time a pair of headphones is put on the variation in placement will alter the frequency response somewhat. That's why professional measurements take an average of at least 3 (or more, Oratory takes 5 as I believe Harman does) re-seats and average them. (I see this as analogous to the listening window averaging being included in the predicted in-room response calculation for loudspeakers in order to account for multiple seating positions / head movement.) This is another area in which HifiMan's headphones seem to do well e.g. here is the frequency response variation of the HE400i with multiple re-seats as measured by Rtings, which combines in-ear mic measurements of several real people wearing the headphones for the bass (including subjects wearing glasses) with HATs measurements in the mids/treble.

I believe useful measurements to base EQ settings on that will work for the majority of users and their specific units and placement on their heads can be achieved, if care is taken to average both several units and placements (as long as variation in both is reasonably low) in order to mitigate all these potential issues.

 

[Maiky76]

It was posted here: https://www.audiosciencereview.com/...eview-closed-back-headphone.19657/post-649070


I don't know though if Maiky's process for computing it is the same as Sean's.

Hi,

Here is my calculation, it is tested against a commercially available analyzer.
I use Matlab.

Pseudo code:
Freq             = 12 points per Octave 20 to 20000Hz reduced to 18000Hz as ASR data stops there…
first Octave: 20Hz 21Hz 22Hz 24Hz 25Hz 27Hz 28Hz 30Hz 32Hz 34Hz 36Hz 38Hz 40Hz

Error            = MeanL/R(FR) - Target ;
STDEV         = std( Error (50Hz : 10000Hz) );
Regression  = polyfit( log( freq(50:10000)), Error(50:10000) ); linear i.e. SLOPE*log(f) + intersect
PPR             = 114.490443008238 -12.6217151040598*STDEV -15.5163857197367*abs(SLOPE of regression)

PPR = Predicted Preference Rating = Score

I have scanned the data from the Harman test:
insignificant difference for scanned data...
STD = 2.02 vs 2.04
Slope = -0.92 vs -0.92
PPR = 74.68 vs 74.38

 

[andreasmaaan]

Close:
1. The Revel F208 is flat anechoically. In a listening room it has this smooth downward tilt. The adjustments of the bass and treble allowed the listener to remove the flattening and get back to the original unmodified target or modify it to taste.

The bass and treble levels were randomly set in each trial, and the listeners' task was to adjust them to taste. Don't get hung up on why it was flattened. Flat was simply one of the settings possible in the listeners' choice of adjustment. It turned out if the speaker is anechoically flat to begin with, little or no adjustment above 300 Hz is needed. That is good news because you can probably equalize your subwoofers and call it a day.

Thanks. I had not realise that the starting shelf settings were randomised, that's helpful.

I'm still a little unclear though on the role of the in-room measurements of the loudspeaker. I do realise the Revel is fairly flat anechoically, and of course that this will not translate into a flat steady-state response in a typical listening room.

However, when you write (my bold):

The headphone was equalized at the DRP to match the flattened in-room steady-state response of a reference loudspeaker (Revel Performa F208)

I take this to mean that the F208 had been equalised to produce a flat steady-state response in the room (which would of course imply that it had been equalised to have a non-flat anechoic response).

Have I understood this correctly? Or have I simply misunderstood that passage from the study?

Apologies for the intensive questions on a very specific technical point.

 

[andreasmaaan]

It is. Jude at Head-Fi has a theory that the extra 4 dB of bass can be explained by measurement artifacts related to these acoustic impedance interaction effects with IEC couplers and IE headphones. If you measure the IE headphones on the B&K 5128 he says they tend to have 3-4 dB less bass closer to the AE/OE target. He has moved to the B&K 5128 because the ear simulator is based on a large study of human subjects and more accurate in terms of acoustic impedance. That may be true but the GRAS people believe the evidence he showed of a Westone IE headphone measured on different IEC couplers was not well controlled. They were measured by different people with different insertion depths, some with ear canal extenders, some with pinna -- all uncontrolled variables that could explain the observed differences. Also the frequency used to normalize the curve was 1 kHz, too high because the couplers have resonators in them that operate down to 800 Hz. The IEC standard recommends normalization at 500 Hz or lower.

We have a B&K 5128 and measured a lot of headphones on all the test rigs. Not had time to analyze the data because of other research priorities.

Ver interesting - thanks again!

 

[amirm]

I have scanned the data from the Harman test:
insignificant difference for scanned data...
STD = 2.02 vs 2.04
Slope = -0.92 vs -0.92
PPR = 74.68 vs 74.38

Oh, that is much better. So what was the "no eq" score then that I linked to earlier?

 

[Maiky76]

Oh, that is much better. So what was the "no eq" score then that I linked to earlier?

Using the scanned data from the Harman test I get the same score than the do
-> that shows my calculation is correct (I had no doubts since Harman is probably using the commercially available analyser I talked you about...)
However with the data you measured the Score is much lower: 63.56 still perfectly correct based on your data
-> (your the sample + your methodology) yield a different score

 

[Sean Olive]

Those measurements on the Drop site were by Jude of Head-Fi. Speaking to Oratory about them he agreed with me that he likely didn't get a good seal when measuring the headphones. Oratory's measurements show better bass extension. As for distortion, it's pretty low across the entire frequency range, not only according to Head-Fi's (90dB@1kHz) but also Innerfidelity's measurements. Subjectively, I don't hear any distortion at all boosting mine up to the Harman target - in fact it's some of the cleanest bass I've heard.



This is an important (and often overlooked) point. Some companies' / models' unit variation does seem better than others though, sometimes belying their price. For example, as @Robbo99999 mentioned, Oratory, who's said most of his posted measurements and EQ profiles are based on multiple units by now, found the HifiMan HE4XX's unit variation to be very low. Direct quote:


That's a $180 planar magnetic headphone (despite planars usually being more expensive than dynamic driver models - I actually got mine for just $140 on promotion, and they've been as low as $120), which has low enough variation between units that individual EQ settings wouldn't even make sense. presumably due to this unit variation being around the same or less than placement variation.

Talking of placement variation, this seems to be another often ill-considered factor when it comes to using headphone measurements for developing EQ profiles, sometimes only a single-placement (and single-unit) measurement being used to base EQ settings on for all users and their units, when in reality every time a pair of headphones is put on the variation in placement will alter the frequency response somewhat. That's why professional measurements take an average of at least 3 (or more, Oratory takes 5 as I believe Harman does) re-seats and average them. (I see this as analogous to the listening window averaging being included in the predicted in-room response calculation for loudspeakers in order to account for multiple seating positions / head movement.) This is another area in which HifiMan's headphones seem to do well e.g. here is the frequency response variation of the HE400i with multiple re-seats as measured by Rtings, which combines in-ear mic measurements of several real people wearing the headphones for the bass (including subjects wearing glasses) with HATs measurements in the mids/treble.

I believe useful measurements to base EQ settings on that will work for the majority of users and their specific units and placement on their heads can be achieved, if care is taken to average both several units and placements (as long as variation in both is reasonably low) in order to mitigate all these potential issues.

What you say makes sense.

 

[Sean Olive]

Thanks @Sean Olive !

One other thing - my experience trying the Harman Curve is full dependant on how accurate orgatory1990's measurements are for my particular headphones...

I am EQ'ing based on his measurement and his recommended adjustments. Any errors in his measurement will affect the EQ to be applied and thus my experience with the Harman Curve.

In other words, I may not actually be hearing the Harman Curve when I think I am !

Agreed.

 

[DDF]

The Harman Target Curve was based on a measurement at the primary listening seat in our listening room using Revel F208 and a JBL M2 calibrated to the Harman Speaker Target.

Thanks for sharing your time on the quorum.

A question about the in ear measurements in room. What length of FFT window is used in these measurements? A longer window pulls in more late room sound and will have a different tilt in the outcome, for the same system. Also, late arrivals have less impact on timbre than the same as an early arrival and so two curves that look the same won't sound the same if one is dominated by more late arrival.

The reason I ask is I think this is an important factor that needs to be considered and consistent when setting targets. The same question applies when measuring a speaker in room and trying to set the window so that non flat off axis can be accurately included in the measure to approximate the perceived timbre. Some argue for variable window (constant 5 cycle vs frequency, etc).

BTW The Goode Beethoven sonata cycle you recommended on your Twitter feed is now my favourite, thanks for sharing that.

 

[Sean Olive]

Thanks for sharing your time on the quorum.

A question about the in ear measurements in room. What length of FFT window is used in these measurements? A longer window pulls in more late room sound and will have a different tilt in the outcome, for the same system. Also, late arrivals have less impact on timbre than the same as an early arrival and so two curves that look the same won't sound the same if one is dominated by more late arrival.

The reason I ask is I think this is an important factor that needs to be considered and consistent when setting targets. The same question applies when measuring a speaker in room and trying to set the window so that non flat off axis can be accurately included in the measure to approximate the perceived timbre. Some argue for variable window (constant 5 cycle vs frequency, etc).

BTW The Goode Beethoven sonata cycle you recommended on your Twitter feed is now my favourite, thanks for sharing that.

Todd Welti did the measurements so I would have to check, but my recollection is the window length is long enough to include direct and reflected sounds.

Agree that the window length is a factor. ON one hand you want it long to get adequate frequency resolution to capture the modal behavior at low frequencies but on the other hand you don't want to capture and equalize interference effects from reflections at higher frequencies. Spatial averaging helps minimize those effects.

Another danger of a long window is capturing the speaker sound power above 500 Hz and equalizing it to be smooth at the expense of making the direct sound worse ( for example filling in a hole at the crossover in the sound power but adding a peak to the direct sound.) That is where a variable window is perhaps useful.

Of course, If the speaker has a smooth or constant directivity, this is less of an issue. There is no substitute for a well designed speaker. Equalization can't fix poor speaker directivity.

Love the Beethoven Sonata Cycle by Goode.

 

[thewas]

This is the actual in-room measurement of the Revel speaker before an after it was equalized to the target curve, We also show the predicted in-room curve from anechoic measurements. It is about a 10 dB slope across the bandwidth.

It's from the referenced paper.


https://www.aes.org/e-lib/browse.cfm?elib=17042

Thank you, except the lower bass it is more around 0.8dB/octave which is my preferred target at many loudspeakers in my current room, I have added it for comparison in pink:

 

[thewas]

We then applied then applied a broadband shelving filter in the bass and treble to adjust the relative bass and treble levels of the speaker (as well as the headphone). These filters could be adjusted to get back to the original in-room target curve that you see in the graph. The point of this exercise was to see if people preferred more bass or treble than the Harman Target Curve or preferred to keep in the same as in a previous study. The same method of adjustment was done to a headphone by moving bass and treble shelving filters up and down.

One discussion point that imho remains though is the way the 2 shelving filters characteristics were chosen (frequency and Q) as those inevitably influence the preferred choice of the gain and also the final target curve. So it would be interesting to hear about more how those were chosen.

 

[andreasmaaan]

One discussion point that imho remains though is the way the 2 shelving filters characteristics were chosen (frequency and Q) as those inevitably influence the preferred choice of the gain and also the final target curve. So it would be interesting to hear about more how those were chosen.

FYI, it's discussed a bit further here (s. 2.7).

 

[andreasmaaan]

A question about the in ear measurements in room. What length of FFT window is used in these measurements? A longer window pulls in more late room sound and will have a different tilt in the outcome, for the same system. Also, late arrivals have less impact on timbre than the same as an early arrival and so two curves that look the same won't sound the same if one is dominated by more late arrival.

This is actually the source of my confusion on this topic. I would have thought that the descriptor "steady-state", in the sections of the study I quoted previously, implied that no window was used (but it appears I may have been mistaken on this point).

 

[Wombat]

It is good to have a considered reference for comparison.

The Harman curve is the result of much research but is by no means compulsory. Preference is allowed. One can see where the preference differs from the Harman curve and match that different response to other products.

 

[blse59]

Did you read the article I linked? I can explain, but doubt I can do a better job than the article.

This a raw response of the ER4S (one of many possible raw responses, each one dependent on measurement conditions):

View attachment 108175

In what respect is this "flat"? And why should it be?

Flat means no boosted bass.

 

[Music1969]

Hi @Sean Olive

What are your thoughts of crossfeed methods for headphones like "Bauer stereophonic-to-binaural DSP":

http://bs2b.sourceforge.net

Have you used it and do you like it?