What Headphone Sound Profiles Do People Like?

Thanks to the decades of well-conducted headphone research from Sean Olive and others, we now have a much better answer to the unsimplified version of the title's question: “What do people typically want their headphones to sound like when reproducing typical stereo recordings?”

But that doesn’t mean we have all the answers. While we as a community have progressed meaningfully in how well we understand & use measurement data, we are still likely farther from fully understanding the answer to this question than we think we are.

Historically, the audio community has struggled with an unfamiliarity with (or outright scorn towards) the research on listener preference, leading many not to engage with measurements at all. Today, however, the problem with the evolving majority is the exact opposite: overconfidence in measurement data.

Driven by an understandable desire for a simple, communicable metric to judge complex subjective experiences, many people take measurements and target curves at face value. Single-line targets are routinely used as blunt pass/fail instruments, and the empirically-focused research outputted by Harman is often overextended far beyond the specific conditions under which it was conducted.

While Harman’s work remains the most well-controlled research ever published on headphone preference, all scientific research has compromises, error, blind spots, and not all research is presented or digested in exactly the same way by everyone.

Which is why in this article, we need to have an honest conversation about how headphone measurements are collected, interpreted, and presented on modern industry-standard fixtures.

First, we’re going to discuss why everyone (including us, up to now) has been using the aggregated output of the Harman research on headphone/IEM preference in an incomplete way that does not sufficiently serve our shared goal of using measurement data to aid in maximizing sound quality.

Then we’re going to share our solutions for this problem: our new Preference Bounds for use with the B&K 5128 & GRAS 43AG/45CA test fixtures, as well as our new measurement database, GraphWorks.

Let’s dive in.

I: The Mechanics & Constraints of Harman’s Work

For anyone who has spent more than a few minutes looking at headphone measurements online, the Harman Target is unavoidable.

“Harman” is brought up in reviews, EQ guides, manufacturer tuning conversations, and is visible on just about every graph comparison tool the community has built for itself.

For most people looking at headphone measurements, their engagement with Harman begins and ends with the Harman 2018 AE/OE target curve, shortened to “The Harman Target” or simply “Harman” in common parlance.

We’ve used it, almost anyone else publishing headphone measurements uses it or has used it, and even when used incorrectly—as most people unintentionally do—it’s such a solid curve that it is still somehow pretty useful.

Most times you’ll see it used to judge a headphone as “good” or “bad,” or as a tool to bring a headphone closer to “good” with equalization. A headphone response that tracks it closely is often assumed to be good, while a headphone that deviates from it is often assumed to be bad.

Fig. 1: An example of a reviewer misusing Harman's work, evaluating frequency responses collected on the HeadAcoustics HMS II.3 against the Harman 2013 AE/OE Target, which was created on a GRAS 43AG-4 with KB007x pinnae.

Audiophiles deeply desire something they can use to indicate or communicate to others that a headphone is “good” or “bad”. Pulling up a measurement & putting it against an authoritative looking target curve is often the easiest way to get an authoritative or convincing answer to this binary question.

But as we’ll get into, the way people draw conclusions based on deviation from or adherence to Harman’s curve(s) was never actually the most robust or most useful part of that research.

Additionally, Harman’s work on listener preference in headphones is not perfect, and it has notable flaws. Some of these flaws are known, some weren’t known until only recently, and some may never be known.

The take here is not that Harman is useless, far from it. It’s best summarized as “Exceedingly few people are able to correctly make use of this well-controlled research.”

So, in the spirit of making the best use of this research, let’s go over some of its issues, so people have the full context for how to avoid misusing this data.

Collection Methodology Issues

In their 2018 over-ear headphone study Harman tested listener preference for 31 headphone responses with 130 trained and untrained listeners⁽¹⁾. This was after almost a decade of doing other headphone listening tests to identify trends of listener preference, so by this point they’d learned a few things about how to design listening experiments to get the least noisy, most useful results.

Controls & Compromises

In 2015 Todd Welti at Harman made modifications to the test fixture they’d been using—a custom set of pinnae (Fig. 2)—for collecting the headphone measurements. They’d identified that test fixtures produced meaningfully different measured bass responses compared to humans measured in equivalent circumstances, so they produced this new ear to make the test fixtures produce a bass response more similar to what humans would get⁽²⁾.

Fig. 2: Todd Welti’s modified IEC 60318-7 pinna, modified to be softer & to make the pinna flange sit closer to the head, injection molded with Shore 00 35 silicone.

In the aforementioned 2018 study, instead of having people try 31 separate headphones to see what sound profiles they liked, which would leave an opening for biases based on non-sound factors to affect results, they opted to “replicate” all of the curves on a single “replicator” headphone⁽¹⁾⁽³⁾. Test subjects wore one headphone and swapped between different sound profiles digitally, which meant biases based on looks, brand, comfort, etc. were eliminated.

Per a comment from one of the researchers in the 2018 study, the researchers ended up choosing a different replicator headphone for this study than in prior tests—opting to use an AKG K712 Pro—because the bosses at Harman were upset that they kept using Sennheiser headphones for their published scientific studies. (Yes, really.)

The researchers also made modifications to the K712 Pro itself to make it seal more consistently to the sides of people’s heads when worn, to ensure ideal bass extension (they used a paint can handle affixed to the existing headband to increase clamp force).

But unfortunately, many of the above choices—some of them made to maximize the utility of this research’s output—ended up having the opposite effect, to varying extents.

Problem 1: The Welti Pinna

In 2015, Todd Welti published his paper “Improved Measurement of Leakage Effects for Circumaural and Supra-aural Headphones”⁽²⁾, where he identified that headphones measured on certain test fixtures either produced:

• More bass than on human heads; because some test fixtures have a perfectly flat, earless surface for the headphone to seal against (eg. flat-plate microphone fixtures)
• Less bass than on human heads; because the GRAS KB0070/0071 (herein referred to as KB007x) artificial pinnae they used at the time were very stiff and non-compliant when trying to seal on-ear headphones.

To combat the latter problem in particular, Welti produced his own silicone ears based on the same standard as the KB007x, but with a few modifications. It was made with a softer silicone, so its deformation characteristics were closer to that of a human ear, which helped specifically on-ear headphones seal on the test fixture.

This new “Welti” pinna was also flattened a bit in its geometry, brought in closer to the sides of the head overall. This again, made it easier for headphones to seal on the test fixture, because they protruded less from the head.

Fig. 3: Frequency response measurements of headphones taken at the blocked ear canal entrance on GRAS KB007x in blue, “Welti” pinna v1 (flattened, but harder silicone) in green, “Welti” pinna Final Version in yellow, flatplate system in red. All traces are compensated to the average of measurements on humans, to show error vs. the human results.

The result is that the discrepancy in bass between humans (0 dB line in Fig. 3) and their new Welti-modified test fixture (yellow line in Fig. 3) was narrowed enough for them to feel confident moving forward with using this ear for their future listening tests, and these tests include the 2018 paper that produced the most popular Harman AE/OE 2018 Target.

This is also where one of the first issues arises. Not necessarily with the research itself, but with how people use it.

While the problem of leakage estimation of test fixtures vs. humans was a smart thing to identify and attempt to control, the new pinna produced meaningfully different treble responses for headphone measurements than the ears people ultimately ended up using when utilizing Harman’s target to evaluate headphone frequency responses.

This wasn’t necessarily the researchers’ fault; this paper was primarily focused on bass, and because it was prior to the popularity of GRAS’s KB5000 series anthropometric pinnae (which is what everyone used after Harman concluded their over-ear headphone research), they simply were not able to get a sense of how the difference in treble between the Welti pinna vs. other pinnae would affect the future use of the research’s output.

But because this treble discrepancy wasn’t really mentioned at length in any of the papers, many of the voices best known for using Harman’s work in our community didn’t know that the Welti pinna was the one used for the 2018 paper, and didn’t know how dramatically different the headphone measurements they produced were.

Which meant for more than half a decade, everyone was actually using the wrong ear with the Harman target… and most of us still are. This matters a ton because the measurements of headphones, and their behavior, is significantly dependent on the ear that it's placed on.

The difference in headphone responses produced by the Welti pinnae and the common, erroneously-used-with-Harman GRAS KB500x pinnae are different enough—in isolation, as well as how they compare to measurements on human ears—that this is not an error that can be ignored.

Fig. 4: The average of 11 headphones measured at the blocked ear canal entrance of GRAS 45CA with KB50xx pinnae (blue), B&K 5128 (green) and GRAS 45CA with Welti pinnae (red). All results are compensated to the average of those same headphones measured on humans at the blocked ear canal entrance (black line), to show the difference between these rigs and the same headphones on humans.(4)

We can see that the two most common fixtures—the GRAS with KB50xx pinnae, and the B&K 5128—overestimate treble above 4 kHz quite dramatically relative to what humans (0 dB line in Fig. 4) are receiving, when measured at the same point. The Welti pinna, by contrast, is much closer to what humans receive in the treble, though it still underestimates treble in certain places.

This magnitude of difference between 5128/KB50xx and Welti means that, for the 2018 target curve made on the Welti ear to be replicated accurately on GRAS KB50xx equipped fixtures or the B&K 5128, we would need to account for the fact that both of these fixtures’ pinnae strongly overestimate the treble above 5 kHz, relative to both the Welti pinna, as well as human ears.

Interestingly, because the Welti ear produces headphone measurements closer to the humans in this dataset, it’s possible that EQing to Harman using headphone measurements done with this rig may have also addressed a commonly-occurring dip in the 6-9 kHz region that tends to pop up on measurements of headphones on human heads⁽⁷⁾⁽⁸⁾.

Fig. 5: Sennheiser HD 650 as measured & Diffuse Field compensated on 69 human subjects at the blocked ear canal entrance (Credit: HUTUBS Database)(7)(8)

This is pure hypothesis on my part, but it is an interesting coincidence that many of the highest rated headphones in this study (including the target) were likely to have the most energy added in this oft-dipped 6-9 kHz band. While further verification is absolutely required to test this hypothesis, comparing this data against human database measurements like HUTUBS above revealed an intriguing correlation.

Regardless, this difference in treble alone between the Welti pinna vs. the others has many implications for responsible use of Harman’s work.

This means the Harman target and its associated model cannot reliably be used with measurements made on ears other than the Welti pinna… and 99.9% of the measurements we see posted in the wild will not have been measured on the Welti pinna.

The overestimation of treble from the aforementioned test fixtures like the publicly-available GRAS or B&K fixtures also has consequences for use of targets that aren’t Harman, as well as for the general interpretation of headphone data and EQ.

We’ll come back to that a bit later.

Problem 2: The Replicator Headphone

In 2013’s Listener Preferences for In-Room Loudspeaker and Headphone Target Responses⁽⁵⁾, Harman made use of their “replicator headphone” methodology for the first time since validating its efficacy(3), and it worked quite well for eliminating non-sound factors from the pool of biases that would influence the preference results they got.

Harman’s researchers used a Sennheiser HD 800 for this study, where they measured it on their test fixture (GRAS 43AG-4 with KB007x ears), equalized it to the in-room response of that same test fixture when measuring a “flat in-room” pair of speakers, and allowed respondents to tilt the response to their preference by controlling the level of bass and treble.

Fig. 6: Frequency response of the Sennheiser HD 800 equalized to the “flat in-room” response of speakers measured on Harman’s test fixture (2013)

The use of the HD 800 for this specific task of being a “replicator” headphone ended up being one of the best choices they could’ve made for this study, but we’ll get to that. For now, here are the totality of the preference results when applied to the HD 800 measured & Diffuse Field compensated on the B&K 5128 and GRAS 43AG-7.

Fig. 7: Adjustments made by subjects in the 2013 Harman paper, but measured, EQed with the same EQ profiles, and Diffuse Field-compensated on the B&K 5128 and the GRAS 43AG-4 with the KB50xx ear, to see what these results would look like ported to newer systems. Results are ½ octave smoothed.

Note: Both of the pinnae used on these fixtures overestimate the treble between 4-12 kHz vs. what a human would receive on average. Diffuse Field compensation, which is applied here for each fixture using its own Diffuse Field HRTF, is not subtracting away these effects because is overestimation is a characteristic of the Headphone-to-ear Transfer Function (HpTF), which is distinct from the far-field transfer function of a Diffuse Field HRTF.

Regardless of what rig we display the spread of preference from this study on, we seem to get a fairly consistent look into the preferences of their test subjects, and this data would likely be replicable by other people using the same type of headphone. We can see a clear broad strokes trend—people preferred more bass and/or less treble than a strictly flat (Diffuse FIeld) response.

This study worked so well, producing results we could replicate within a reasonable margin on newer systems, because the Sennheiser HD 800 is a uniquely good headphone for this task. Due to the HD 800’s acoustic design, it has benefits that make it especially well-suited to being a replicator headphone:

• It is more tolerant of small leaks, due to its extremely open acoustic front volume
• It is quite free of pinna deformation, due to its large earcups
  • The above two qualities also make it less positionally-variant
• It has a low acoustic output impedance; changes in input load have lesser effects on measured response than we’d see on headphones with a higher acoustic impedance
• Quality control & manufacturing consistency of the HD 800 is close to the best you can get in the retail headphone market

Unfortunately, the AKG K712 Pro was not good for any of the above things.

• The K712 Pro has an open-ish front volume, but it is likely still of a quite high overall acoustic impedance, evidenced by it being rather leak intolerant in our internal measurements
• While the K712’s earcups are wide in circumference, they’re not especially deep, so pinna deformation is a potential concern for some ear sizes
  • Because it doesn’t have great performance in the above two factors, it's likely also quite positionally variant when worn by humans (again, corroborated by our internal measurements)
• Quality control is meaningfully worse than the HD 800, both in terms of channel matching as well as unit-to-unit consistency (mostly matters for end-users wanting to use this research for corrective EQ)

This is where we must remember: Any of the curves that have been produced by Harman’s work are, in actuality, a frequency response measurement of a headphone measured on a test fixture, with some smoothing applied.

This means if the headphone’s frequency response is unstable, prone to variation based on either positioning, pinna deformation, or the acoustic load of the system it's placed on (whether its leaky, or if the ear is presenting a different input impedance), we cannot make perfectly predictable assumptions about the frequency response people would have actually received.

Furthermore, since most headphone-to-ear transfer functions produce narrow-bandwidth features in the treble, this means that what is EQed to look smooth on one test fixture may well be very unsmooth on another. Remember this, it’ll come up later.

The important thing here is that AKG K712 Pro is a perfect example of a headphone with an unstable frequency response, one that is eminently mutable based on the ear it is measured on and how it is placed.

This means we cannot extend assumptions of target deviation & adherence vs. a single headphone’s EQed and smoothed frequency response to other ears, because as we’ve talked about multiple times on our platform, headphone behavior varies based on the ear it is placed on. This fact is true of all headphones, including the Sennheiser HD 800, but it is especially true of the AKG K712 Pro.

Now, as mentioned earlier, the researchers themselves actually identified that this was a flaw with the AKG K712 Pro, which is why they modified its headband with a paint can handle to increase the clamp force⁽¹⁾. This meant the listening test subjects were more consistently able to get a seal when wearing the headphone.

But while this was almost certainly the best call for controlling this variable in this study (short of choosing a better headphone), it still serves as a bit of an issue for those who want to use this research outside of the conditions it was conducted… even though Olive et. al explicitly warned us not to do this with their research from the very beginning:

“...we caution the reader not to generalize or extrapolate the conclusions of these experiments to conditions outside those that were tested.⁽⁵⁾“

In the simplest terms, any effort to transpose the results of Harman’s 2018 listening tests must necessarily account for both the specific frequency response idiosyncrasies of the AKG K712 Pro, as measured on the Welti ear, while also accounting for the instability inherent to this headphone’s frequency response across different ears.

These are meaningful issues that need to be addressed, but unfortunately the data collection methodology is not the only component that makes it a bit tough for us to properly understand & use what is still the most well-controlled research we have.

Data Analysis Issues

A later 2019 paper from the same group at Harman used hierarchical clustering on the preference results from the 2018 paper to identify and segment out “classes” of listeners with similar patterns of ratings. The authors then took the five most highly rated headphones in each class, averaged their frequency responses, and used those averages to describe the sound profile associated with each class.⁽⁶⁾

This segmentation paper serves as the analysis of the 2018 listening tests, and it’s often reduced to a neat summary: 64% of listeners prefer Harman, 15% prefer more bass, and 21% prefer less bass. It is an understandable shorthand, but it is actually not a totally accurate description.

For example, when it comes to Class 1 (“Harman Lovers”) and Class 2 (“More Bass enjoyers”), Harman was actually the #1 most preferred target response for both groups.

So saying “15% prefer more bass,” isn’t exactly true when this class still significantly preferred the Harman target (while also liking 2 or 3 other headphones that had more bass than the Harman target).

Additionally, while the target was obviously the highest-rated response on average, it was also not always clearly separated from every other response in every test run. It certainly was not the case that people thought Harman sounded good, but everything else sounded bad.

For example, while the target was ranked towards the top in every aggregated panel, it was statistically tied with HP19 (Shure SRH-840) in Test Four and with HP25 (Philips Fidelio X1) in Test Five.

Fig. 8: Averages & 95% confidence intervals of listener preference ratings for Test Four & Five from A Statistical Model that Predicts Listeners’ Preference Ratings of Around-Ear and On-Ear Headphones (2018)

This gives us some evidence that while the target was a very strong performer, it was not the only response listeners enjoyed.

Furthermore, when comparing all of the subject’s responses across tests without anchor-based rescaling, there were a good few other headphones that scored well, quite closely to the Harman Target, despite not being the #1 choice.

Tipping the Scales

Test subjects were able to rate the virtual headphones on a 100-point preference scale, but they generally avoided the extreme ends of that scale. The Harman target averaged between 63.5 and 71.2 across the five tests, while the low anchor (Meze 99 Classics) averaged between 17.1 and 20.5.

This left the researchers with a 100 point scale where only ~54 points were being used. Rather than treat the test subjects’ use of the scale as the final preference scale, they opted to linearly expand the ratings to occupy a full 0–100 range between Low Anchor and Target, respectively.

The authors described this as a correction for contraction bias, and that is a reasonable thing to try and account for. People are often hesitant to use the very top or bottom of a rating scale, particularly in a comparative listening test where the listener is asked to judge multiple responses together.

While the Target & Low Anchor scored near or at the top and bottom respectively in all cases, their reported means and rankings still varied slightly from panel to panel. Some headphones scored higher than the Target, or lower than the Low Anchor. This meant the scores were free to travel over 100 or under 0 into the negatives after the results were re-scaled.

And again, this is where more issues arise. While scale normalization & anchor-based rescaling are standard statistical procedures used to correct for scale-use variability across test blocks in studies like this, the issue was not that they normalized for contraction bias, but that their specific choice to use linear anchor-stretching distorted the relative preference distances between top-performing headphones.

In other words, their rescaling exaggerated the degree of difference between top-performing headphone curves, making alternatives to the target look worse on paper than they were in the actual listening test results.

To restore the true relative preference distances, we reverted the anchor-stretching math to display the raw, un-rescaled preference ratings for Class 1, keeping the relative placement of the preference results intact, and what we get is this:

Fig. 9: Un-rescaled average listener preference scores from Class 1 of Segmentation of Listeners Based on Their Preferred Headphone Sound Quality Profiles (2019)

Harman was still the best curve tested on average, but right below it there are headphones that still tested rather well compared to Harman, and more importantly, compared to the rest of the headphone responses tested in this study. This group of “enjoyable, but maybe not as enjoyable as Harman, on average” headphones is also quite clearly separated from the next lowest group of headphones.

Class 1 gave the Harman Target a mean preference rating of 70. It also gave six non-Harman Target response profiles quite high ratings. While the original paper states that four headphone responses were equally preferred to the Target, we would say that these seven responses were closely grouped enough to be treated as “similarly good to one another”:

• Harman Target: 70
• Sony MDR-7506 - 67
• Shure SRH-840 - 66
• Oppo PM-1 - 66
• Philips Fidelio X1 - 65
• Sennheiser HD 800S - 63
• Beyerdynamic DT 990 - 63

And if you actually look at these curves, while there are certainly differences, the overall trend is fairly obvious:

Fig. 10: The Harman 2018 AE/OE Target, with the six other similarly preferable curves.

While there are things in here that look meaningfully different than the Harman target, notably the Oppo PM-1 with its flat bass response & reduced 3-6 kHz region, the overall trend is that things that share the hallmark features of Harman’s target (HRTF-like shape, with more bass and less treble) tend to be quite well received by listeners.

These headphone curves are close enough to the target—both in score, and in frequency response trend—that it feels imprudent to use this fairly wide grouping of profiles to describe these listeners as simply “The 64% that prefer the Harman Target.”

Even if they like Harman, they like other things too, and this is the case with almost any listener. Nobody enjoys only one frequency response every time they’re tested, and Harman’s own research reflects this too.

Fig. 11: Oppo PM-1 vs. the Harman target, illustrating that two extremely different frequency responses can be similarly preferred by the same people

Now, this isn’t meant to imply that every listener would be equally happy with every curve. But we do think that because even starkly different curves, like the Oppo PM-1, are within such a small distance score-wise, it makes sense to re-examine how we as a community think about this research.

Because the dominant class was not narrowly attached to one response, and most listeners aren’t either, we need to leave space in our discourse and interpretation of measurement data that accommodates more than one single target line as “good”.

Not only that, but this work also gives an inkling as to what sort of “non-Harman” response curves may be a defensible thing to aim for if you’re trying to maximize sound quality for a large segment of the population!

Before we finish up this recounting of Harman’s work though, we need to talk about the AKG K712 Pro one last time.

More Than Just Headphones

As we touched on in Problem 2: The Replicator Headphone, headphone targets from this era are almost always measured frequency responses of headphones, with smoothing applied.

But headphone measurements aren’t just headphone measurements: they’re measurements of the full acoustic system composed of the headphone and the ear it’s placed on. Which means the headphone and the ear both need to be known/quantified before making any assessments of how that measurement will relate to an actual listener.

This means that prior efforts to replicate the sound of this 2018 Harman AE/OE target using different replicator headphones on different fixtures were starting from a faulty premise⁽¹⁰⁾. Because the interaction of headphone geometric & load behavior is headphone-specific and ear-specific, any attempt to replicate the results of this study would absolutely need to start with the AKG K712 Pro.

So let’s consider the AKG K712 Pro (smoothed to 1/6th octave) measured on the Welti ear versus the Harman 2018 target.

Fig. 12: Sean Olive’s original measurement of the AKG K712 Pro from A Statistical Model that Predicts Listeners’ Preference Ratings of Around-Ear and On-Ear Headphones (2018), plotted against the Harman AE/OE 2018 Target

Obviously, the K712 Pro is not very adherent to the Harman target, so to get this headphone to replicate the response of the Harman target, corrections needed to be made.

For this headphone to match the target, the sub-bass needs to be lifted, the low midrange needs to be reduced, the upper midrange needs serious reshaping— the ~1 kHz and ~3 kHz areas need a lift, while 2 kHz needs a slight reduction—the low treble and mid treble needs a lift, especially around 8 kHz, while the upper treble needs to be cut around 10 kHz, etc.

Fig. 13: Sean Olive’s original measurement of the AKG K712 Pro (orange) and the error curve relative to the Harman target used for equalization, from A Statistical Model that Predicts Listeners’ Preference Ratings of Around-Ear and On-Ear Headphones (2018)

This is where the nature of making narrow-bandwidth adjustments to a “smooth” target on a test fixture comes back to bite us. Even if these narrow-band adjustments are not what we’d ordinarily recommend for correcting headphones, that is what was used in the original study. In other words: The Harman 2018 AE/OE headphone target is the sound of the specific AKG K712 Pro used in this study, with the EQ profile above (dotted blue line) applied to it.

This means even the narrow band, ear-specific and headphone-specific adjustments—like the 8 kHz dip being filled in, the 11 kHz peak being notched out, the 13 kHz being boosted dramatically—are still adjustments that reflect exactly what people heard when listening to this target response and reporting their enjoyment.

Because this headphone—and every other headphone in the study—has specific features when measured on the Welti pinna, these features affect what frequency response features were ultimately heard by the subjects in the listening panel.

But since different ears commit different treble response filtering on input, the frequency response measured on other ears is likely to look a little weird, since doing so is essentially replicating the inverse of the ear-specific effects from the original ear that aren’t likely to show up on the second ear.

For example, when you apply the same EQ to the same type of headphone, an AKG K712 Pro with an extra clamp modification, and measure it on a different fixture like the B&K 5128, you get quite a different result:

Fig. 14: AKG K712 with a clamp modification measured on the B&K 5128 with the original filtering as derived from the measurements on the Welti pinna equipped fixture to make it adhere to the Harman 2018 AE/OE Target

And as an example of the interactions being, well, headphone-specific, if we do the same thing for a different headphone—the Sennheiser HD 800, also measured by Sean Olive on the Welti pinnae-equipped fixture—we get a different result:

Fig. 15: AKG K712 with a clamp modification and Sennheiser HD 800 measured on the B&K 5128 with filtering derived from their measurements on the Welti pinna-equipped fixture to make them both adhere to the Harman 2018 AE/OE Target

Early on in the time of 5128’s adoption, many people in the headphone science community tried doing something like this. However, because our understanding of inter-fixture relatability at the time was flawed/incomplete, most thought this to be something that obviously can’t represent the sound of the actual Harman target. After all, look at those massive treble elevations! Where is the inviting visual smoothness of the Harman target we love so dearly?!

This denial was based on erroneous ways of thinking about measurements and targets. As said, headphone targets up to this point had predominantly been headphone measurements with generous smoothing applied, but if you move a headphone that is EQed smooth on one ear, to a different ear, it will be producing meaningfully different high-Q features in the treble that will cause what was “smooth” on one ear to look quite “unsmooth” on another, even if they sound the exact same to an actual listening test subject.

The fact remains: the above measurement of the AKG K712 Pro with extra clamp, EQed with the same filters from the original study, measured on the 5128, is likely the closest we’ll ever be able to get to transposing the 2018 Harman target to the B&K 5128, as it is the most direct method possible to do this.

But now that we know that the target was not the only likable response for the dominant class of people, and we know that different headphones EQed to the same target even on the original measurement rig will produce different results when measured on another rig, it’s worth asking:

Do we stop there, and consider this effort satisfactory enough to tell the whole story?

We think not.

II: Introducing Our New Preference Bounds

Before diving in I want to mention that from here on out we’ll mostly be looking at Diffuse Field compensated data. I’ve done a whole explainer of what this means and why we use it, but I’ll briefly summarize before we go deeper.

We’ve used, and will continue to use Diffuse Field compensation with our Preference Bounds because it is a necessary acoustic baseline to understand & calibrate away the average HRTF features of the head and ears measuring the headphones/IEMs, so we can focus on the behavior of the headphone(s) without having to account for the frequency response of the ear itself.

Diffuse Field is used here not as a standalone target curve, but as a standardized, fixture-specific compensation baseline that allows us to compare headphone behavior from different heads/test fixtures without having isolated ear-specific features muddying up the results.

Because Diffuse Field on its own is absent an empirically-backed preference element, we’ve always had some preference element overlaid on top of it, whether it’s been a -1 dB/octave downslope, or a set of Preference Bounds to serve as the empirically-backed preference envelope. We will continue to do this.

Moving on.

Headphones

In 2023 we introduced our original Preference Bounds, and the goal was simple: we wanted something from Harman’s work that allowed us to indicate to viewers that the truth of headphone preference is not that a single line is good and everything else is bad, because Harman’s own work controverts this notion.

Unfortunately, only Harman’s 2013 paper “Listener Preferences for In-Room Loudspeaker and Headphone Target Responses”(5) gave us sufficiently granular listener response data to draw a probability space based on the respondents “average” preferred adjustments, as well as their upper/lower confidence intervals.

The goal of our old Preference Bounds was never to replace the Harman target with a more complicated target, and it was also never meant to predict an individual listener’s preference from a graph (just as Harman’s model cannot perfectly predict a headphone’s preferability from a graph).

While we don’t want to add confusion to a discourse that is already deeply confused, we also want to make sure we’re reflecting the truth of headphone preference as it’s been exhibited in every one of Harman’s studies comparing headphone curves; listeners rarely only like one thing.

The goal with our new Preference Bounds is instead to take the listener preference data Harman collected seriously enough to preserve its context, and give people a full understanding of the title. We want to show our audience “what headphone sound profiles people like,” in the format they’ve come to expect from us.

That means treating the target as one preferred profile among a small set of other profiles that the dominant class also rated highly.

So we recreated all of the well-performing profiles—the Harman Target, Sony MDR-7506, Shure SRH-840, Oppo PM-1, Philips Fidelio X1, Sennheiser HD 800, and Beyerdynamic DT 990—but we did it in a way that we feel is more defensible than prior efforts.

We measured the clamp-modified AKG K712 Pro on the GRAS 43AG-7 with KB500x ear and the B&K 5128, adjusted them using the same exact EQ profiles applied to the AKG K712 that would’ve been used to simulate the other curves in the original 2018 study, and then compensated all of the responses with each fixture’s Diffuse Field HRTF.

Some may argue that porting the 2018 filter transfer functions onto modern 5128 and GRAS fixtures is an interim step until new double-blind tests are conducted natively on these rigs, and we agree. However, laying a static Welti target curve directly over a KB500x or 5128 measurement without accounting for pinna transfer functions is simply incorrect. Synthesizing the underlying 2018 filter transfer functions onto the ear geometries/impedances that are actually producing most of the data in our space is the most scientifically rigorous bridge available today if we want to use this research (and of course, we do).

The below graph shows the actual field of preference variation among these “good” curves, including the Harman target, across both of the industry-standard measurement fixtures we use.

Fig. 16: Measurements of the AKG K712 Pro measured on GRAS 43AG-7 (blue) and B&K 5128 (green) replicating the frequency response EQ curves for the Harman 2018 AE/OE Target, Sony MDR-7506, Shure SRH-840, Oppo PM-1, Philips Fidelio X1, Sennheiser HD 800, and Beyerdynamic DT 990. All curves are compensated with the Diffuse Field HRTF of their respective test fixture, and progressively-smoothed with a self-developed smoothing algorithm (1-octave smoothing until 2 kHz, to 1/3 octave at 3 kHz, 1/6 octave at 4 kHz, continuing to 1/12 octave at 10 kHz.)

And what we get may look surprising, but remember what we went over in the prior sections:

• These rigs overestimate the response in the 4-12 kHz band, which means it likely shouldn’t look like the oft-discussed “linear downslope” used as a proxy for the broad conclusions of Harman’s work (ie. “people prefer less treble and more bass than speakers that measure flat in a room”).
• While the field of results here very clearly has a strong 8 kHz peak and 10 kHz dip, this is a result of the K712 Pro measured on the Welti ear producing a strong 8 kHz dip and 10 kHz peak when measured on these specific fixtures. Regardless of how it looks, this is the best characterization of what the listening test subjects ultimately received and preferred would look like on these systems.
• The results between the two fixtures do not perfectly agree, because they have different geometry—which affects the results above 3 kHz quite strongly, even after Diffuse Field compensation—and because they have different input impedances, which has effects across the entirety of the frequency response.

To turn this field of curves into a properly usable Preference Bounds space, we reject the outermost 10th percentage of results to eliminate any obvious outliers, while also making a judgment call similar to the original research not to put too much stock into fine-grained features above ~10 kHz or so (especially those that show up in only one headphone response’s measured result).

And the result is this, our new Preference Bounds for ITU T P.58 Type 3.3 (GRAS) and Type 4.3 (B&K 5128) fixtures, meant solely to be used with Diffuse Field compensated measurements of over-ear headphones:

Fig. 17: Headphones.com’s new Preference Bounds for ITU T P.58 Type 3.3 (GRAS) and Type 4.3 (B&K 5128) fixtures, meant solely to be used with Diffuse Field compensated measurements of over-ear headphones

The above may look strange at first glance (particularly in the treble), but this field of variation contains the best possible transposition of the totality of the preferable results from Harman’s 2018 paper to the fixtures we and our community use, and we hope just looking at this provides people with context they have been missing from prior visualizations of Harman’s output for years now.

Keen readers may have already expected it to look like this, because for a long while now we’ve been looking at Diffuse Field compensated headphones measured on, for example, the B&K 5128, that have shown behavior in the treble that seemed at odds with what people were actually hearing.

And there's a good reason for that: As mentioned prior, Diffuse Field compensation alone does not account for the overestimation of treble when these rigs are measuring headphones, because the closely-coupled earfield interaction of a headphone is both of a different directionality of source and of a meaningfully higher collective acoustic impedance than the far-field interaction of measuring speakers around the head in a Diffuse Field.

This disconnect, as well as the Headphone-to-ear Transfer Function on the ears of these fixtures overestimating treble, is why you'll often see a large 8-9 kHz peak even on headphones that (can) actually dip strongly in this region on human heads, like the Sennheiser HD 650 we looked at human measurements for earlier.

Fig. 18: Sennheiser HD 650 measured & Diffuse Field compensated on the B&K 5128, displayed against a -1 dB/octave downslope

Our 2023 Preference Bounds directly lifted the method-of-adjustment filters from Harman’s 2013 paper. While faithful to that specific experiment, that study's adjustments were constrained by a fixed second-order 105 Hz bass shelf. This created a pronounced 200 Hz 'bass tuck' in the Preference Bounds that overestimates how problematic this area of the response would actually be for people. 

In faithfully reproducing the treble adjustments solely through their adjustment values (using a first-order high shelf), the exact shape of the frequency response is also not quite as accurate as it could be for actual headphones measured on the systems we use (as seen in Fig. 7).

Fig. 19: Sennheiser HD 650 measured & Diffuse Field compensated on the B&K 5128, displayed against our old Preference Bounds

Over-indexing for this specific shape of low-shelf boost can make HD 650 seem too “warm” in the midrange, despite many enthusiasts and recording professionals finding it one of the most accurately-produced midrange tunings in the entirety of the headphone sphere. Over-simplifying the frequency response in the treble to simple shelf filter shapes underestimates the magnitude and complex shaping of the treble that headphones measured on our fixtures would actually produce.

So unlike this older approach, our new Preference Bounds now account for both the idiosyncratic treble behavior of headphones measured on the rigs people currently use, while also accommodating a wider breadth of preferable bass elevation shapes.

And as a result, the HD 650’s compliance to what most would call an “excellent-sounding midrange and treble presentation” is much more obviously laid out, while aspects most people would object to like its bass rolloff or its mid-treble being a bit on the relaxed side, are shown much more clearly.

Fig. 20: Sennheiser HD 650 measured & Diffuse Field compensated on the B&K 5128, displayed against our new Preference Bounds

For these reasons, we think our new Preference Bounds are a serious leap ahead of prior visualizations of headphone measurement data done on ITU T P.58 Type 3.3 and 4.3 systems, including our older Preference Bounds, as well as the Harman Target itself.

This does implicitly mean that the “Diffuse Field + flat slope” method that has been used with over-ear headphones up until now by many (especially on sites like squig.link) should probably fall by the wayside, at least when it comes to measurements on these fixtures. The flat slope was only ever a crude estimation of the general trend of bass-to-treble delta that people in the speaker and headphone literature prefer, but as we've talked about, people tend to take anything that looks like a target and treat it like one, even when it shouldn't. 

So I'll say it here again: A simple flat slope does not account for the overestimation of treble that these modern industry-standard test fixtures actually have, so the "flat slope targets" people are using should probably be eschewed going forward.

Just as single-line targets fail over-ear headphones due to earfield interactions between the headphone and fixture, in-ear monitors face the exact same issues, but amplified further by the complete absence of pinna interaction, as well as variations in ear canal length & volume.

Which leads us to the next big update: Our new IEM Preference Bounds.

IEMs (BETA)

We did in fact make new preference bounds for IEMs measured on the B&K 5128 based on Harman/Sean Olive’s work from 2016 to 2025, but there are a few things people need to know about this data before we go further.

The data we’re using is based on Sean Olive’s measurement of the Sennheiser Momentum IEi taken from the Brüel & Kjaer Type 5128. IEMs measured in this fixture tend to portray two artifacts that don’t show up in human measurements:

• There is a “rocking mode” between 100-200 Hz that arises due to a spring mass resonance being formed between the tension of the eartip in the silicone ear canal entrance of the 5128 + the mass of the IEM itself. This resonance gets excited by sine sweeps (the stimulus we use to measure IEMs) and the result is a little Z-shaped or N-shaped artifact that will be present around 200 Hz in some of the curves shown
• There is a slight (~1 dB) underestimation of the 400 Hz band for IEMs relative to what humans receive when wearing/listening to the same IEMs.⁽⁹⁾

With these idiosyncrasies in mind, let’s move to the data.

For the goal of making a boundary space to describe the range of listener preference in IEMs, we adopted a similar approach as before. We took Sean Olive’s measurement of the Sennheiser Momentum iei, applied all of the EQ profiles preferred in Harman’s aggregated work, and then used Diffuse Field compensation (using the Headphones.com IEM Diffuse Field HRTF) to see a full spread of what listeners actually prefer.

The below field of variation includes the totality of the preferred responses from Harman/Sean Olive’s work between 2016 and 2025, including of course the Harman IE 2019 Target and the SoundGuys target.

Fig. 21: All preference adjustment results from Harman’s IEM studies 2016-2025, as replicated on a Sennheiser Momentum iei measured on B&K 5128

Adding a boundary to encapsulate this field of variation (which excludes the two 5128-specific artifacts previously mentioned), we get our BETA Preference Bounds for IEMs measured on ITU T P.58 Type 4.3 systems like the B&K 5128.

Fig. 22: Headphones.com’s BETA IEM Preference Bounds for in-ear monitors measured on ITU T P.58 Type 4.3 Ear Simulators (B&K 5128)

Again, this might look a bit odd to people in the treble, but this is where, like headphones, IEMs have narrow-bandwidth deviations based on positioning, anatomy, and impedance, which mean what you see on a measurement may in fact sound surprisingly different to any given listener.

Because the research into IEM listener preference is relatively much less settled than with over-ear headphones (much smaller sample sizes in the listening test, and results that at times disagree with conclusions established in earlier research), and because we’re not 100% confident in what data to exclude to eliminate outliers, we’re releasing these bounds in Beta form, as a field that includes all preferred results, with the expectation that they will likely change in the future as more research becomes available.

Applying Preference Bounds derived from over-ear headphones to in-ear monitors as we've been doing up to now was always an imperfect, interim compromise necessitated by the lack of consolidated in-ear preference literature done on the newer fixtures. Now that we have that data and more, we get a glimpse into how acoustic coupling, the lack of pinna interaction, and ear canal volume/length variations contribute to slightly different preference results for in-ear monitors vs. over-ear headphones.

As an example of how these bounds differ from the old ones, let’s take a look at how one of the most well-liked IEMs out there looks against the old bounds vs. the new bounds.

Fig. 23: Subtonic Storm measured on B&K 5128 and compensated using Headphones.com’s IEM Diffuse Field HRTF, displayed along with our 2023 Preference Bounds

Compared to the old bounds, one would feel pretty confident in saying that this IEM, the Subtonic Storm, is likely to be too bright for most people starting around 5 kHz, because it’s right up to the edge of the boundary space there.

Additionally, one might say (based on the measurement alone) that Storm’s 2.7 kHz area may be a little too relaxed for the majority, because it’s actually veering outside the bounds slightly in that region.

However, fewer people actually have these complaints about the Subtonic Storm than this measurement visualization may suggest. Even though some people with darker-than-average HRTFs like myself, or people with long ear canals and HRTFs with a relatively relaxed 3-5 kHz region like Resolve find Storm a bit too intense around 5 kHz, many people really seem to love this IEM.

And when you look at it against the newer Preference Bounds, it’s pretty clear why:

Fig. 24: Subtonic Storm measured on B&K 5128 compensated using Headphones.com’s IEM Diffuse Field HRTF, displayed along with our new BETA Preference Bounds for IEMs

The added range in the lower-bound of the bass frequencies more firmly cements Storm as a “bass boosted” set relative to the range of potential preference in the literature, but this part, as well as the midrange, still look excellent and well within bounds.

But the 2.7 kHz dip is now properly contextualized as something maybe slightly relaxed, but unlikely to be a massive problem to most people, while the 5 kHz region actually looks like it’s almost exactly where a majority of people would want it to be.

While yes, Storm still flirts with the upper end of the Preference Bounds in the upper treble, it also gets near the top of the bounds in the sub-bass, which most people would say is a fairly common approach to balancing colorations when it comes to IEMs (“V-shaped”).

And here is my current favorite IEM (Prisma Lumen) vs. the new Preference Bounds:

Fig. 25: Prisma Lumen measured on B&K 5128 compensated with Headphones.com IEM Diffuse Field and displayed along with our new BETA Preference Bounds for IEMs

We can see that yes, it is still firmly a bass-boosted set, but the midrange and treble compliance is excellent across the board such that what people actually receive when listening is likely to be quite preferable, barring any anatomical variation in the treble (which will still mean that some people who hear Lumen will find it either too shouty, too relaxed, too dark, or too bright).

Known Limitations

The end result here is not that we now know within the space of a single boundary what “good” looks like, nor does this explicitly tell us what “bad” looks like.

Because with a field of variation that’s this wide, it is simply not possible to use it as a way to disqualify bad products; even responses like flat, unadjusted Diffuse Field would fall inside this boundary, even though it has been losing in listening tests since the early 2010s.

This is a big part of why this boundary space is currently in BETA, instead of being a finished product. We’re not exactly sure how to best limit the bounds in certain regions to avoid allowing for obviously unsafe responses, because many of the respondents in the 2025 study that provide much of this data ended up preferring a very reserved (some even near-zero) bass elevation, along with boosted treble.

However, even if this boundary space isn’t super useful for disqualifying obviously bad products, we as a community need to ask ourselves: have measurements under the prior paradigm really been useful for that up to this point?

If you go to a random squig.link site or AudioScienceReview, you’ll find measurements almost-exclusively plotted against single-line targets that, regardless of the actual intention of the operator who collected the measurements, gives a very clear, authoritative message to the reader: If it matches the target its good, and if it deviates, that’s probably bad.

But this has neither actually been the case—false negatives are myriad with this evaluative process—nor has it been supported by most of the research that produced these curves to begin with.

Which brings us to the question many might be asking: What’s the best way to use these new bounds correctly?

III: Introducing GraphWorks

When we introduced our initial Preference Bounds in 2023, those early bounds propagated across disparate web-based graph engines, and they were frequently used without consistent documentation, rig-specific Diffuse Field HRTF calibration, or guidelines on how that boundary space should be interpreted.

While web-based tools like squig.link deserve credit for democratizing measurement data across the hobby, without unified operator standards or rig-specific compensation profiles, low-quality uncompensated data leads to widespread misinterpretation by viewers and reviewers alike.

To address this, we created GraphWorks—a dedicated measurement repository built specifically around our modern visualization standards for our data collected on the ITU-T P.58 Type 3.3 (GRAS) and Type 4.3 (B&K 5128) systems.

Rather than reducing complex acoustic interactions to a single pass/fail line, GraphWorks is designed to display properly collected and compensated headphone and IEM measurements within the Preference Bounds.

GraphWorks is available today as an open utility for reviewers, manufacturers, and enthusiasts seeking to evaluate headphones and IEMs through a scientifically grounded, pluralistic lens, and I can’t wait to see how people make use of it, as well as other novel tools & features we have planned for it.

But anyway, let’s go ahead and wrap this up.

Conclusion

The goal of this article has been to illustrate to people exactly why strict adherence to the Harman target is not only impossible for most people to actually get, but an exercise in futility when squaring the actual truth of Harman’s collective work. People need to know the sources of variability and uncertainty that affect the headphone measurements we take for granted, as well as reckon with the plurality of listener preference that Harman’s work observed and reported.

Harman’s work has flaws that need to be acknowledged so we can make the best possible use of their well-controlled, consistently predictive. These flaws mean that their most popular target response is neither easy to replicate, nor a sufficient condensation of the entirety of the conclusions that should be gleaned from their seminal work.

Headphones themselves are not static in their performance, their sound is fluid based on the ear and head they are placed on, and the perception of that sound is largely influenced by the individual anatomy and preference of the person listening to that headphone.

We need to accept that single-line targets were never a sufficiently good tool for understanding or communicating the above, nor were they ever as good an answer to the question “What headphone sound profiles do people like?” as they could’ve been.

Hopefully this piece brings to light enough of the context and acknowledgment that people have been missing when it comes to Harman’s work and the broader sphere of headphone measurements done on industry standard fixtures, such that readers can walk away less confident, but more informed than they were prior. Thanks for reading.

This article would not have been possible without considerable input from Blaine LaCross, who served as the primary thrust behind scraping all of the data from Harman’s papers, as well as the continuous support from Headphones.com that allows pursuit of educational audio endeavors like this. He has a fantastic seminar from CanJam NYC 2025 where he goes over a lot of the contents of this work. I would like to thank them both, as well as Konstantin Davy, Joel Merrifield, Sean Olive, Todd Welti, Elisabeth McMullin, Omid Khonsaripour, Fabian Brinkmann, Etienne Rivet, Andrei Bezzola, Gunther Thiele, Dorte Hammershøi, Henrik Møller, and everyone else who has contributed substantially to my ability to put together this article. 

Appendix

(1) S. Olive, T. Welti, O. Khonsaripour. 2018. A Statistical Model that Predicts Listeners’ Preference Ratings of Around-Ear and On-Ear Headphones. In: Proceedings of the Paper 9919; AES Convention 144; May 2018: Audio Engineering Society. p. 9919:. https://aes.org/publications/elibrary-page/?id=19436

(2) Todd Welti “Improved Measurement of Leakage Effects for Circum-Aural and Supra-Aural Headphones.” Journal of The Audio Engineering Society (2015): https://www.aes.org/e-lib/browse.cfm?elib=17699

(3) S. Olive, T. Welti, and E. McMullin, "A virtual headphone listening test methodology." Proceedings of the Audio Engineering Society 51st International Conference, Helsinki, Finland, August 21–23, 2013.

(4) S. Olive "Will the Headphone Industry Ever Agree on a Design Target Curve?" Danish Sound Day 2022. Online. Available: https://danishsoundcluster.dk/danish-sound-day-2022-a-great-networking-day/

(5) S. Olive, T. Welti, and E. McMullin, “Listener Preferences for In-Room Loudspeaker and Headphone Target Responses,” in Proc. AES Convention 135, Oct. 2013, Paper 8994. Online. Available: https://aes.org/publications/elibrary-page/?id=17042

(6) S. Olive, T. Welti, and O. Khonsaripour, “Segmentation of Listeners Based on Their Preferred Headphone Sound Quality Profiles,” in Proc. AES Convention 146, Mar. 2019, Paper 10156. Online. Available: https://aes.org/publications/elibrary-page/?id=20289

(7) F. Brinkmann, M. Dinakaran, R. Pelzer, P. Grosche, D. Voss, and Stefan Weinzierl (2019). “A Cross-validated database of measured and simulated HRTFs including 3D head meshes and anthropometric features.” J. Audio Eng. Soc.

(8) Fabian Brinkmann, Manoj Dinakran, Robert Pelzer, Jan Joschka Wohlgemuth, Fabian Seipl, and Stefan Weinzierl The HUTUBS HRTF database. DOI: 10. 14279/depositonce-8487

(9) Rivet, Etienne & Smeaton, Calum & Kim, Jeong-Woo & Welti, Todd & Olive, Sean & Baker, Jeff. (2025). Estimating the sound pressure at the eardrum and ear canal transfer function with in-ear headphones. :https://www.researchgate.net/publication/395586371_Estimating_the_sound_pressure_at_the_eardrum_and_ear_canal_transfer_function_with_in-ear_headphones

(10) S. Olive and D. Clark, “A Preferred Headphone Target Curve Defined for the B&K 5128 and GRAS45CA-10 Test Fixtures,” in Proc. Express Paper, Oct. 2025, Paper 394. Online. Available: https://aes.org/publications/elibrary-page/?id=23068