New Compensation Curve for InnerFidelity Measurements

Nice surprise ending to this one.

You may remember my trip to Harman International last year to plonk my measurement head down in front of a reference level speaker system. My motive was to determine exactly what speakers looked like with my particular Head Acoustics dummy head. My basic assumption was that good headphones should measure like good speakers. If I could capture the head response with good speakers, I could then build a compensation curve for headphone measurements that would be related directly to my particular head and a speaker based reference.

Now, is there a solid, scientific reason a compensation curve should be developed like this? No, not really. As many of you know, developing a headphone target response curve or compensation curve is still in its infancy. Sean Olive at Harman International has been working hard on the problem, which I've reported on here and here. He did find during his research that using an approximated in-room speaker response for headphones was preferred.

However, when I tried applying the Harman Target Response curve directly to my measurements of a number common headphone it didn't fair very well. That's when I got really motivated to develop a compensation curve for my head in particular.

If your interested in the story in detail, check out the articles here, here, and here. To recap: I took measurements of Harmans reference system with my head at five azimuth (left to right) angles (-20, -10, 0, 10, and 20 degrees) with the head in three elevation angles (-10, 0, and +10 degrees). I did this series of 15 measurements with both speakers playing and with only the left channel playing. Here's a recap of the measurements:

Here's the in-room response measured with a regular measurement microphone. 170913_Blog_IFCompensation_Graph_InRoomResponse

You'll notice two things about this curve: a 5dB bass boost below 150Hz; and a gentle, overall warm tilt. Simply put, this curve is a result of putting speakers that are flat in an anechoic chamber into a room. The roughly 1dB/octave tilt come from reductions in sound power in the room as higher frequencies from the speakers narrow and beam. Warren TenBrook wrote a great article about this curve and more, which includes some interview materials with Sean Olive. Highly recommended.

Immediately after taking the measurements I summed up and averaged all the stereo measurements for a first look. 170913_Blog_IFCompensation_Graph_FirstPassStereoTotal

Right off the bat, I'll draw your attention to two things. First is the bass boost. It looks too big to me, and it's got a big hump centered at about 40Hz. This doesn't look like the room response; I've got problems with it.

The other thing to note is the peak at 3.5kHz and the two little sombrero bumps to either side at 900Hz and 12kHz. As you'll see, the measurements have a lot of noise in them, but this sombrero feature continues to show up.

I then crunched some numbers to come up with some graph families of the data taken of the response with the left speaker only. These curves were smoothed in Excel with some simple moving average filters.




Note that there's quite a bit of activity in the sombrero areas to either side of the 3.5kHz peak. It seems the area between 800Hz and 2kHz is somewhat more sensitive to elevation changes, and the area between 8kHz and 15kHz is more sensitive to azimuth change.

Thinking my simple smoothing might be creating artifacts, I got in contact with Arnaud, who crunched the numbers with more sophisticated filters. Here's the family plots in that form.





And then looking at the family averages.



And then comparing the average of stereo vs. left channel only measurements.


It's pretty obvious this filtering ended up poorly representing the bass response, but it did do a pretty good job of providing a cleaner look above 1kHz.

Then I spent a lot of time just staring at the plots.

In the end I came to the conclusion that trying to play any games like tuning the curve according to the lower elevation plots to bring the sound down in your head, or trying to tune out narrow peaks, is rife with problems. Simple is better, in my view. Basically, I ended up just averaging all the curves together and smoothing it a bit.

Another thing I decided is to simply use a flat bass. There's plenty of controversy around the bass hump in the target curve, even I feel like I'd prefer the bass hump to have a different shape. Bottom line: I think it would be least confusing to simply leave the bass flat and let the viewer interpret it as they wish.

Here's my new compensation curve (in gray) compared to the diffuse field and independant of direction compensation curve.


Here are some plots of well known headphones using the new compensation curve. I'll show the before plot using the independant of direction compensation currently used on spreadsheet, and then the after plot using the new compensation curve.

Sennheiser HD 600 Before

Sennheiser HD 600 After

Sennheiser HD 800 Before

Sennheiser HD 800 After

Audeze LCD-4 Before

Audeze LCD-4 After

Mr. Speakers Aeon Before

Mr. Speakers Aeon After

NAD VISO HP50 Before


I'd love to hear your thoughts on this new compensation in the comments below!

And now for the surprise!
You may be wondering at this point, why the heck am I doing this? Am I going to go back and poke these new compensation curves into the 800+ measurements spreadsheets I have and republish them? No thank you, that's way too much work, and all these .pdf spreadsheets are a pain in the butt to maintain.

No, I'm doing it because the InnerFidelity Graph Tool is in the works! Code is being written as you read this! Wa-hoooo! I can't wait for us all to be able to compare headphones live on the web again. And I think developing a good compensation curve is the most important thing I can do before the roll-out.

Over the coming few months I'll tell you more about the graph tool as things get firmed up. I'll certainly be giving a hard look to what's been measured and what needs to get measured to have as complete as possible database of measured headphones at launch time. InnerFidleity readers will, as always, have the opportunity to send in headphones missing from the database—InnerFidelity will pay shipping both ways in the continental US.

More soon!

Puffy's picture

I'm curious why you think that the bass hump isn't some resonant artifact of the measurement head that needs to be addressed in the compensation curve.

Here's my theory: It's due to the way that the sound waves impact the measurement head causing the air inside of it to resonate resulting in higher sound pressure on the measurement mic.

It is definitely centered around a harmonic of the big hump at ~3.5kHz.

I'm prolly just pulling things out of my ass though. I haven't reviewed my physics for some time now.

Iliketrains's picture

I think because addressed in previous posts, that people do like the elevated bass. And if you are using the same measuring head, maybe you would want to get the same "artifact"to replicate the same resonances when comparing.

Tonmeister's picture

The bass bump (4 to 6.5 dB depending on age/training of listener) is based on some research we did where listeners adjusted the level of bass (and treble) of an anechoically flat loudspeaker in our reference listening room. The steady-state in-room response of an anechoically flat loudspeaker usually has a downward tilting curve (-10 dB tilt from 20-20kHz) which represents a combination of the sound power and the direct sound at high frequencies. The exact curve depends on the absorption characteristics of the room, but for average domestic rooms this is a good statistical average.

Tonmeister's picture

Here are two references related to preferred bass and treble levels of headphones and loudspeakers:

zobel's picture

The best looking graph is the HD600 ...left channel. That wasn't the pair that ate your screwdriver on one of the drivers, was it?

zobel's picture

it was a pair of needle nose pliers that slammed into the driver's magnet, mashing that diaphragm flat! I'm guessing that is why that channel measures down 8 db at 7 and 14 khz, along with the boogered up dropped off bass on that side. Sennheisers are never that whacked out on their own.

Pokemonn's picture

Is web graph tool underdeveloping? Hallelujah!

Argyris's picture

The HD 600 and HP50 graphs match up pretty well with what I heard (and still hear with the HD 600) from both headphones, much better than the previous compensation curve. I felt the HP50 had too much mid- and upper bass* and a bothersome spike in the mid-treble, with a sort of well in between cutting out some of the upper harmonics on vocals. The HD 600 fares much better and, apart from the weak sub bass, its plot looks almost ideal.

*Though, strangely, distinctly poor extension--I suspect the pads weren't sealing on the bottom despite the vice-like clamping force.

speakerguy1's picture

Just out of curiosity were any of dummy head measurements of the speakers made with one ear of the head pointed towards the speakers? Since the dummy head is designed for what you could call on axis measurements it seems like it would be worth trying in several orientations to confirm the shape of the ear is not affecting these in room measurements.

andy133's picture

Just wondering why we think the compensation curve should be subtracted from the measurement? Do we have a good argument why the offset is independent of the sound level? Has anyone checked whether the compensation curve has the same shape with sweeps at various levels? Non linearities can creep everywhere.

GNagus's picture

I've been doing this for a few years with Innerfidelity, by placing each headphone's responses in a tab, and clicking back and forth between the tabs.

brkitup's picture

Thank you so much for your work on this. The graphs have traditionally not represented very well what the end user hears, and the new graphs are much more representative of how these headphones sound, especially in the treble. The one thing that still bothers me slightly, though, is the dip around 4-5kHz that seems common to all headphones when measured and compensated. I'm not sure how accurate that part of the compensation is.

Argyris's picture

Same here. It's smaller in width and depth in the IF compensation than on the DF or ID compensations, but it's still there. I wanted to see what this sounded like filled in, so I tried it on my HD 600 with a parametric EQ. I dialed in a simple peak centered around 5 kHz, with 0.3 octave width and 5 dB boost--it doesn't perfectly fill the hole, but it's a quick and dirty test of concept. It didn't sound horrible, like filling in the much larger and deeper trough left by the other compensations does, but it did add an uncomfortable edge and glare to everything. Of course my HD 600 might be different than the measured one, but one thing Sennheiser is known for is the consistency of its performance across units. I can't assume without proof that my particular unit is atypical enough to be grossly boosted in this area compared to a "normal" HD 600, such that adding more energy puts the output above baseline.

It could be an issue with translating the curve to headphones; there might be some as-yet undiscovered factor that complicates things (like an ear canal resonance; IEMs, incidentally, almost universally have a big spike right in this area). It could also just be preference; after all, the data were collected by placing a dummy head in a room with what Harman (and Tyll) consider to be good speakers. Maybe both like a bit of extra zing around 5 kHz?

zobel's picture

The curve we have now is not what we hear, but what we measure in a dummy head at eardrum location. It may represent what would be measured in a human head there as well, but that isn't what the brain 'hears', any more than a camera in a human eye would record what we see. The eye and ear are not the end of the system in real life. The brain assembles the data into a 'corrected curve' to represent reality as it is, not strictly from the nerve impulses.
Mr Speakers asks the same old question that we all have been for many years, "Why not portray flat as a flat line, using raw data to provide the difference curve between measurements, and the perception of flat. It of course is not possible to do this scientifically or subjectively, but that attempt needs to be made to provide a more meaningful and representative curve. It would be an averaged curve, like Harmon's, based on listener's trials with a reference speaker system in a room, followed by headphone listening to determine the closest match to the room sound.

zobel's picture

The raw measurements below 1 kHz represent a subjectively accurate image of what we hear without modification.

HuoYuanjia's picture

Please just provide the RAW measurements in a way that they are easy to read. Having 10 thick grey lines over another makes it very hard to distinguish them. Please have them averaged or select the two most reasonable.

me klasse's picture

Hey Tyll,

I think the new compensation looks better so congrats for that.
The new graphs fall a bit to quickly starting at around 700Hz up to 1500Hz, not sure if I can hear that. Otherwise the treble looks better. I've always thought HD800's brightness was misrepresented on your spreadsheet.

Here's the question:
(Since the perceived tonal balanced is volume dependent - more or less, along the equal loudness contours)
What's the volume @1kHz for which (you think) the new compensation is targeted?

On a side note, sad to hear about the spam filter problem. I was writing for a while the other day. Thanks for letting us know what happened anyway.

Me x3

lerrens's picture

Innerfidelity has the best headphone measurement database right now and the Graph Tool will make it super awesome! I'm so excited I have to create an account just to tell you this: Thanks for all your work! Really looking forward to this :)

castleofargh's picture

I don't get it. we start with a clear hypothesis about having the sound of speakers in a good room and go for it. then you don't like the results, which is fine, it's one idea with one room an one dummy head. then we enter the twilight zone.

if your real purpose is to have a compensation curve that reflects your impressions, maybe you could just draw it by hand for several headphones, make compensations out of it based on raw, and then average that.
but taking real data and manipulating it until it looks the way you feel, I don't like it one bit. only ugly marketing should do that sort of stuff.

and TBH it annoyed me when you made your own Harman curve too. it works, it doesn't. you have all the necessary data or not, that's life. but cherry picking the objective elements you subjectively like, that's not objective anything, but it's like alcohol free beer, it really looks like beer.
I kind of like the final curve, I'm sure you get how that's not my issue.

now that I've lectured you like I knew anything, it's the best time to ask for something(I'm such a great negotiator). I'd like to see the raw front looking dummy(no angle), with stereo and with only one speaker as you seem to have both. to see how much the response changes on that head from correct "crossfeed"(IDK how that should be called)of both channels, against one channel for one ear like headphones.

and congrats for the graph tool being made. that's really good news.

geniekid's picture

I kind of agree here. It feels like you just gave up trying to calculate the compensation curve at the low end because of an anomaly that couldn't be explained, although I suppose it doesn't matter as long as you're using the graphs for comparing different headphones.

viscoelasticity's picture

When measuring with a dummy head in a room, diffraction, refraction and reflection of the sound field around the head will create a spatial acoustic shadow zone. This phenomena then effects the measurements at the ear positions (leading to the Sombrero hump), but would not influence the single microphone measurement (see Harmon's room measurement above).

The cut-on frequency varies from 3-6 KHz, and is often found at 3.5 KHz, as noted above. This is relatively constant due to the geometry of the head, as noted in Tyll's summary of the measurements.

Does it then make any sense to try and use this type of dummy head room measurement to develop a headphone target curve, where the sound field is immune to head diffraction effects?

JMB's picture

In addition to the room influences on the measurements dummy heads are made of different materials than real heads. I was reading that bone has a very unique combination of high stiffness and high vibration damping and only some exotic alloys including In can approach or better that and both parameters influence the acoustic properties of that structure.
Likely the material which is used to make the pinna will have different properties (including damping and absorption) than real ones. Obviously shape variations (especially of the targus) will have acoustic influences.

mrspeakers's picture

Interesting work, Tyll, thanks again for your contributions to the community!

One thing I note is that the curve appears to create a notch/discontinuity around 1K. I suspect it very unlikely almost all these headphones have a midrange discontinuity like that, it's a pretty critical area, and so it looks to me like there's an issue there that might create concerns about midrange linearity.

At the broader level, I've always had the concern no compensation curve actually allows a headphone with a reasonably flat acoustic response (assessed via sweeps, noise, etc) to display as "flat." The general form of the curve above 1K probably makes intuitive sense to enthusiasts with regular exposure to headphone response curves, but I think they'd be quite counterintuitive to people not as steeped in the craft.

In the absence of a standard "flat" compensation curve what about publishing a difference curve relative to your target? Musing out loud, people would be able to easily see where the response is on target or off and thus even if they're not able to intuitively process what the 4K dip really sounds would a differential plot would make it more intuitive?

zobel's picture

I thank you for this post. That has been my exact question for years. We need a flat graph to represent flat cans, not a graph of the effects of ear geometry. The difference curve you mention would be extremely helpful and actually meaningful in representing what we hear, not just what we measure.

detlev24's picture

Thank you for your comment. :)

A flat/neutral frequency response should not be the aim for any headphone; it not even is the target for loudspeakers in the actual listening environment. Loudspeakers, amongst other important characteristics, need to measure flat on-axis in an anechoic chamber and that is it!

While every room is different, loudspeakers - to achieve natural/accurate reproduction - need to be individually calibrated, but that is only the final and simplest part of the whole setup process (e.g., using JBL's Intonato 24).

Over-Ear headphones come with a pre-defined "room" - finalized by the earpads - and thus the final target response curve should be achieved by design, rather than DSP. If we measure sound waves directly before hitting the [outer] ear; that will be the same for every individual. Only thereafter, individual anatomy changes the game and it is impossible to take account of that. Consequently, would Over-Ear headphones mimic the acoustic space emitted by calibrated loudspeakers [see at Harman] around our ears - that could be the answer to natural headphone sound. L/R loudspeaker channels treated separately; and if wanted, good crossfeed could later be achieved by devices like the RME ADI-2 Pro or SPL Phonitor x.

As far as I know, Sonarworks try to do exactly this during their calibration process. While not perfect, it translates better than anything else to what I hear on my JBL M2. It certainly would be interesting to measure one of their individually calibrated headphones pre- and post-; and to compare it to Tyll's target. Anyways, no personal preference of any reviewer should translate into measurements!

Of course, this only would work very limited for on-ear headphones and not at all for in-ear monitors. Especially for the latter, only calibration considering individual anatomy would really provide an accurate listening experience.

wktenbrook's picture

Tyll and Dr. Olive, check out Figure 1 in this post:

The JBL M2 curve that Charles Sprinkle loaded into the DSP prior to your measurements is very close to the Reference Room speaker target "RR". These two sister target curves have the highest level of bass, consistent with your head measurements.

I wrote:

"...note that Figure 4 in Olive-Welti-McMullin AES paper 8867 (2013) presented headphone listening tests where corresponding curve RR1_G was preferred over curve RR_G. Presentation slides summarizing the preferences were posted by Dr. Olive.

In my opinion, the evidence weighs in favor of curve RR1 as a neutral reference curve over the typical Reference Room house curve, RR, used during Tyll's measurements..."

So, speaker curve "RR1" has 2-4 dB less bass 100 Hz and below and was documented as preferred by Harman when used as a basis for the headphone target.

I think your measurements bore witness to how the fingerprint of the bassier selected speaker compensation curve can be seen throughout the process.

I don't have any insights into the high frequency response, other than our discussion of comb filtering when two speakers played versus none with a single left speaker.

csglinux's picture

Great work - and I love the idea of the comparative online tool!

I seem to have missed the point of the new calibration curve though. If the new curve is supposed to be more accurate, why would we want to smooth away all the detail (and completely remove any low-frequency effects)?

I'm used to looking at Innerfidelity's compensated curve plots as is. I guess my brain doesn't want to re-learn new curve shapes without understanding what's behind this "progress". Being able to plot deltas will be awesome, but as an earlier poster mentioned, there's often an awful lot of variance between measurements - often larger than difference amongst headphones. See for example the Monster Turbines, vs the Monster Turbine Copper Pro. I'd swear they were the exact same headphone. On the other hand, look at the IE800 - it's L and R measurements seem to have come from two totally different headphones!

One final point on IEMs. We really all (myself included) have to get into the habit of stating what tips they were measured with. Tips have a huge effect on the spectra, especially foam vs silicone.

zobel's picture

Nope. You cant measure what we see by putting a camera in an eyeball, any more than you can measure what we hear by putting a microphone in an ear.

Subjectivity rears its beautiful head here. The brain subjectively interprets the SPL data at the eardrum, and trough the entire inner ear as it is converted into a representation of SPL in nerve impulses. It wouldn't help to measure those impulses on the way to the brain either since the processor (brain) actually synthesizes our picture of the sound, with it's own correction curve, which is as individual as our set of ears are.

We will just have to make tests like Harmon did to find averages among listeners in comparing cans to a reference system.

francisk's picture

I'm really looking forward to the InnerFidelity Graph Tool to be able to compare headphones live on the web. I'm feeling like a kid again.

Jazz Casual's picture

I refer to Tyll's headphone measurements as an objective data point, but it appears that he's just arbitrarily gone and shifted the goal posts.

markus's picture

I wonder how the frequency response would change if the dummy had arms and an upper body.

tony's picture

We've been marketed the concept that "Ruler Flat" is desirable, for a good many Decades.

It's nonsense

Real World is a Curve.

Back in 2015, after not being able to hear the supposed wonderfulness of the Schiit Yggy DAC, at the Ann Arbor headphone meet, I had my own hearing evaluated at the University of Michigan's Ear Specialists.

Not only was my hearing well off "Flat", my headphone system ( Schiit Asgard 2 & Sennheisers ) was well off the ideal "Curve" .

Thru the magic of Equalization, UofM adjusted my little system to closely match Harmon's Curve, taking my own hearing ability into account.

Viola, I have a wonderful music system that I'm not in any hurry to replace or up-grade ( although I'm planning a Focal headphone and perhaps an Feliks amp.)

Tyll & Bob Katz are the leading folks on all this for the entire Audiophile World. I can't find any other Consumer based reviewers mentioning anything about any of this.

For me,

Tyll is Professor Tyll!

Bob Katz is a visiting Professor

Tony in Michigan

tinyman392's picture

If you had a copy of the plain text numeric data for your graphs, you could probably write a quick script that would generate all the PDFs again for you in seconds.

zobel's picture

The raw data as measured trough Tyll's dummy head represents perceived sound below say, 900 Hz? Yes it does, so we can leave it alone when trying to conger a subjective response curve. I think we do appreciate and desire the bass lift of a good listening room. When your head IS the listening room, boosting the bass to make headphone "match" room measurements producing a measured curve to approximate the preferred room make the cans WAY too bassy. I am fairly certain the RAW data below 900 Hz is a good approximation of what we subjectively consider accurate.