The original paper recommended tone controls. The ideal situation would be Harman compliment FR (though due to technical limitations, the actual best would likely be harman but without a bass shelf) and tone controls to allow people to dial in the bass and treble to individual preference. You can also find tune from there. My preference for bass and treble can even vary from song to song due to the mix.Interestingly, the headphones with the Harman curve sound harsh to my ears around 5kHz, just as test results in 5128.
-
Welcome to ASR. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!
AES 2025 Paper: New targets for the B&K 5128 GRAS 45CA-10
- Thread starter amirm
- Start date
DualTriode
Addicted to Fun and Learning
- Joined
- Oct 24, 2019
- Messages
- 993
- Likes
- 659
The human part of this headphone testing paper is the most challenging part.
Now that Dr. Olive has retired and is no longer at Harman he does not have the resources for a larger pool of human subjects.
This paper used only 13 Dan Clark employees as listeners to "validate" the test results.
I am not rushing out to purchase a B&K 5128.
Now that Dr. Olive has retired and is no longer at Harman he does not have the resources for a larger pool of human subjects.
This paper used only 13 Dan Clark employees as listeners to "validate" the test results.
I am not rushing out to purchase a B&K 5128.
Last edited:
This was a very interesting project, and while I feel it answered some questions, it also raised a few.
The following are the complete set of "Target Curves" for both couplers, an explanation of how the curves are derived and why they are different follows. The best way to understand the Target Response curves is to realize that that is what the measured frequency response of the headphone should be after EQ is applied to sound the same as the original Welti-Target. The variance in response by headphones shows how variable different targets are across the GRAS and 5128 systems, the greater the variability, the harder it is to create a general target.
The syntax is the 5128 and KB5000 curves were derived from JUST the NOIRE X measured on the Welti, 5128, and KB5000 setups. The _7 measurements were a group of 7 headphones measured on the 3 systems (5 planar magnetic, 1 electrostatic, 1 dynamic, 5 open 2 closed measured by me), and the _20 were 20 headphones measured previously by Sean's associate Omid (80% closed).
Since the way these curves are generated is complex, I thought it might be worth reviewing what you're actually seeing. The "Target Curves" Amir showed were derived by taking the Measured Response (or the individual or averaged headphones)/Welti Target Response to create a difference function. This difference function was conjugated with the original Welti Curve to create a "Target" based on the difference for that headphone or average for the KB5000 and 5128.
Difference function = (individual Or averaged headphone frequency response)/(Welti Harman Target Frequency Response)
Target = Difference Function conjugate Welti Curve
From this you can see why we have multiple curves to test, as the difference function varied widely based on the individual or averaged responses. Hence the listening tests to determine which came closest to target, and why the "target curves" in fact look different.
In effect, if every headphone had the same difference on each fixture relative to the original Welti target every curve would be the same. The fact that they differ is indicative of the fact that each coupler has its own unique impedance, as does each target headphone, or average of a group of headphones. The upshot and key conclusion is that the average target is NOT exact, though the individual headphone measurements are if the headphone impedance is extremely low, for example the Corina.
Between the _7 and _20 curves of average responses the KB5000_7 curve can be considered a validated target based on listening test resuts, but it was interesting that the original KB5000_20 curve was different and performed less well, and neither the 5128_7 or 5128_20 passed statistical muster. This is likely driven by the fact that the KB5000_20 and 5128_7 curves were based on 20 headphones, 80% of which were closed (increasing effect of DUT acoustic impedance), while the KB5000_7 and 528_7 curves had 5 open headphones, one stat, three planar, and a dynamic, and 2 closed headphones, and probably had less variance due to acoustic impedance. However, neither averaged group of headphones on the 5128 passed listening tests, showing the inherent variability of measurements on this fixture and the effects of acoustic impedance could not be controlled to the extent the resulting target was acoustically close to the original Welti target.
Interestingly, however, if you refer to the headphone "Target" curves in Amir's first post (shown below), you'll note that there was not only the expected variability in the lower frequencies but that some open headphones exhibited even greater variability in the higher frequencies across fixtures, while other open and closed had appreciably lower variance. This compound variability at both ends of the spectrum made the averaged 5128 curves less accurate, and accounts for why the single headphone derived curves performed better.
In the end, I think these two graphs if standard deviation of targets for the 5128_7 and KB5000_7, and why the 5128_7 underperformed in listening tests.
Figure 3: Standard deviation of target curves for 5128 and KB5000
Assessing Target Response and Acoustic Impedance of Measured Headphones
When looking at the "Target response" curves, additional details become obvious.
First and foremost is that some headphones show more correlation between adjusted target curves than others, in particular the Corina, an electrostatic with very low acoustic impedance, has the most consistent target profiles for all three couplers, and could be considered a benchmark for mapping the "ideal" headphone across all three fixtures (in terms of ability to provide a consistent target and resulting sound).
Somewhat to our surprise, the other 4 open headphones all showed greater variation across the 3 fixtures than did Corina, or the 2 closed back planar headphones. Also, the Susvara and 518 had more variance in bass response than Stealth in the low frequency response. We lacked an acoustic impedance fixture to assess DUT Z across the full spectrum, so it is possible, especially for the 5128 where greater variance is seen, that this is both a combination of ability to seal on the 5128 "head" vs the flat plates of the Welti and KB5000 setup, but the conventional wisdom is that lower acoustic impedance is assumed for open headphones, and it is assumed that this should reduce the effect of poor seal, so this explanation seems inadequate.
Also, seal does not explain the significant high frequency variance of the HiFiman and Sennheiser open-back headphones. More investigation is required to fully explain this.
Questions for future investigation
1) How does acoustic impedance of the DUT vary by closed and open topologies, and within each topology what accounts for the large variations seen between a very low acoustic Z design and open designs that indicate greater variability? Are these variations consistent with conventional wisdom that open headphones are lower Z than closed? What is the frequency dependence of this? Are different target curves recommended based on DUT impedance?
2) Would the coupler setup of the 5128 be more consistent when used with a flat-plate design, allowing for a more predictable and accurate target response?
I will note as a designer that we rely heavily on the 5128 when tuning our AMTS because we have found in our listening tests that variations in response above 3K on the 5128 correlate very closely to what we hear, more so than the GRAS.
Also, we have found numerous examples of instances where the GRAS shows a large response dip while the 5128 shows a sharp peak at the same frequency. There's a lot to unpack in this...
For those interested in using the KB5000_7 curve to assess headphones and to make compensation suggestions, a couple of recommendations we'd make to people are:
1) If you develop filters to "fit to curve" using the KB5000_7 target (attached below) don't use High-Q filters, especially above about 2.5KHz. Above that level we suggest 1/2 or 1/3 octave smoothing, as variability in ear acoustic impedance will materially skew alignment between your ear and the coupler, potentially exaggerating problems.
2) Above 5-6KHz consider simply using a simple shelf unless there are glaring "features" to correct. This recommendation is consistent with the original Harman research.
A final note; later this year DCA will begin to publish results and PEQ settings using corrected KB5000 curves, and possibly for headphone-specific curves using the 5128 (so long as we have the original Welti ears on hand to generate exact individual headphone targets to create valid EQ results), as we feel this will provide people with the best fit-to-cuve experience. We'll probably also include these in the OPRA system.
KB5000_7 Target for Audio Precision (edit header and footer to use with REW), allows normalization at any frequency
The following are the complete set of "Target Curves" for both couplers, an explanation of how the curves are derived and why they are different follows. The best way to understand the Target Response curves is to realize that that is what the measured frequency response of the headphone should be after EQ is applied to sound the same as the original Welti-Target. The variance in response by headphones shows how variable different targets are across the GRAS and 5128 systems, the greater the variability, the harder it is to create a general target.
The syntax is the 5128 and KB5000 curves were derived from JUST the NOIRE X measured on the Welti, 5128, and KB5000 setups. The _7 measurements were a group of 7 headphones measured on the 3 systems (5 planar magnetic, 1 electrostatic, 1 dynamic, 5 open 2 closed measured by me), and the _20 were 20 headphones measured previously by Sean's associate Omid (80% closed).
Since the way these curves are generated is complex, I thought it might be worth reviewing what you're actually seeing. The "Target Curves" Amir showed were derived by taking the Measured Response (or the individual or averaged headphones)/Welti Target Response to create a difference function. This difference function was conjugated with the original Welti Curve to create a "Target" based on the difference for that headphone or average for the KB5000 and 5128.
Difference function = (individual Or averaged headphone frequency response)/(Welti Harman Target Frequency Response)
Target = Difference Function conjugate Welti Curve
From this you can see why we have multiple curves to test, as the difference function varied widely based on the individual or averaged responses. Hence the listening tests to determine which came closest to target, and why the "target curves" in fact look different.
In effect, if every headphone had the same difference on each fixture relative to the original Welti target every curve would be the same. The fact that they differ is indicative of the fact that each coupler has its own unique impedance, as does each target headphone, or average of a group of headphones. The upshot and key conclusion is that the average target is NOT exact, though the individual headphone measurements are if the headphone impedance is extremely low, for example the Corina.
Figure 1, Target Curves for 5128: Red=Welti, Orange=NOIRE X 5128, Purple=5128_20, Blue=5128_7
Figure 2, Target Curves for KB5000: Red=Welti, Orange=NOIRE X KB5000, Purple=KB5000_20, Blue=KB5000_7
Figure 2, Target Curves for KB5000: Red=Welti, Orange=NOIRE X KB5000, Purple=KB5000_20, Blue=KB5000_7
Between the _7 and _20 curves of average responses the KB5000_7 curve can be considered a validated target based on listening test resuts, but it was interesting that the original KB5000_20 curve was different and performed less well, and neither the 5128_7 or 5128_20 passed statistical muster. This is likely driven by the fact that the KB5000_20 and 5128_7 curves were based on 20 headphones, 80% of which were closed (increasing effect of DUT acoustic impedance), while the KB5000_7 and 528_7 curves had 5 open headphones, one stat, three planar, and a dynamic, and 2 closed headphones, and probably had less variance due to acoustic impedance. However, neither averaged group of headphones on the 5128 passed listening tests, showing the inherent variability of measurements on this fixture and the effects of acoustic impedance could not be controlled to the extent the resulting target was acoustically close to the original Welti target.
Interestingly, however, if you refer to the headphone "Target" curves in Amir's first post (shown below), you'll note that there was not only the expected variability in the lower frequencies but that some open headphones exhibited even greater variability in the higher frequencies across fixtures, while other open and closed had appreciably lower variance. This compound variability at both ends of the spectrum made the averaged 5128 curves less accurate, and accounts for why the single headphone derived curves performed better.
In the end, I think these two graphs if standard deviation of targets for the 5128_7 and KB5000_7, and why the 5128_7 underperformed in listening tests.
Figure 3: Standard deviation of target curves for 5128 and KB5000
Assessing Target Response and Acoustic Impedance of Measured Headphones
When looking at the "Target response" curves, additional details become obvious.
First and foremost is that some headphones show more correlation between adjusted target curves than others, in particular the Corina, an electrostatic with very low acoustic impedance, has the most consistent target profiles for all three couplers, and could be considered a benchmark for mapping the "ideal" headphone across all three fixtures (in terms of ability to provide a consistent target and resulting sound).
Somewhat to our surprise, the other 4 open headphones all showed greater variation across the 3 fixtures than did Corina, or the 2 closed back planar headphones. Also, the Susvara and 518 had more variance in bass response than Stealth in the low frequency response. We lacked an acoustic impedance fixture to assess DUT Z across the full spectrum, so it is possible, especially for the 5128 where greater variance is seen, that this is both a combination of ability to seal on the 5128 "head" vs the flat plates of the Welti and KB5000 setup, but the conventional wisdom is that lower acoustic impedance is assumed for open headphones, and it is assumed that this should reduce the effect of poor seal, so this explanation seems inadequate.
Also, seal does not explain the significant high frequency variance of the HiFiman and Sennheiser open-back headphones. More investigation is required to fully explain this.
Questions for future investigation
1) How does acoustic impedance of the DUT vary by closed and open topologies, and within each topology what accounts for the large variations seen between a very low acoustic Z design and open designs that indicate greater variability? Are these variations consistent with conventional wisdom that open headphones are lower Z than closed? What is the frequency dependence of this? Are different target curves recommended based on DUT impedance?
2) Would the coupler setup of the 5128 be more consistent when used with a flat-plate design, allowing for a more predictable and accurate target response?
I will note as a designer that we rely heavily on the 5128 when tuning our AMTS because we have found in our listening tests that variations in response above 3K on the 5128 correlate very closely to what we hear, more so than the GRAS.
Also, we have found numerous examples of instances where the GRAS shows a large response dip while the 5128 shows a sharp peak at the same frequency. There's a lot to unpack in this...
For those interested in using the KB5000_7 curve to assess headphones and to make compensation suggestions, a couple of recommendations we'd make to people are:
1) If you develop filters to "fit to curve" using the KB5000_7 target (attached below) don't use High-Q filters, especially above about 2.5KHz. Above that level we suggest 1/2 or 1/3 octave smoothing, as variability in ear acoustic impedance will materially skew alignment between your ear and the coupler, potentially exaggerating problems.
2) Above 5-6KHz consider simply using a simple shelf unless there are glaring "features" to correct. This recommendation is consistent with the original Harman research.
A final note; later this year DCA will begin to publish results and PEQ settings using corrected KB5000 curves, and possibly for headphone-specific curves using the 5128 (so long as we have the original Welti ears on hand to generate exact individual headphone targets to create valid EQ results), as we feel this will provide people with the best fit-to-cuve experience. We'll probably also include these in the OPRA system.
KB5000_7 Target for Audio Precision (edit header and footer to use with REW), allows normalization at any frequency
Attachments
This is an amazing explanation thank you @Dan Clark
Was this done with the original modified GRAS 45CA setup or just the KB5000 and 5128? So would I be right in thinking that, if there is no individual reference GRAS 45CA for each headphone, the graphs only show the difference of the combined effects of headphone impedance and coupler impedance? For example, if you point to the varying elevation at 8kHz in the 5128 measurement for some headphones (Stealth, Noire XO, Susvara) which does not appear at all in others (Corina, Noire X), you cannot say definitively this is due to the headphone or the coupler? I think I am being slow and this may be what you are getting at in you're 1st question for future investigation.
Was this done with the original modified GRAS 45CA setup or just the KB5000 and 5128? So would I be right in thinking that, if there is no individual reference GRAS 45CA for each headphone, the graphs only show the difference of the combined effects of headphone impedance and coupler impedance? For example, if you point to the varying elevation at 8kHz in the 5128 measurement for some headphones (Stealth, Noire XO, Susvara) which does not appear at all in others (Corina, Noire X), you cannot say definitively this is due to the headphone or the coupler? I think I am being slow and this may be what you are getting at in you're 1st question for future investigation.
For this to be accurate we had to recreate the original “Welti” setup, we had a pair of the Welti ears and used the original RA045 mic, then compared to the KB5000 ears with the RA0402 and 5128. That gave us the correct baseline measurement to use when generating the Difference Function that created the Target Response curves you mention. Without the Welti ears and the original mic you couldn’t recreate the Harman curve baseline, which was essential to mapping the curve to the system people actually have.
This is the most complex point to get across, but another way to say it is we measure the headphone on the original welti setup, then on the new target. We calculate the difference in the response on those two fixtures then adjust the original target to a new target based on that difference. If all headphones had zero acoustic impedance the adjusted targets would have all been the same.
So if the mapping works correctly, the Target Curve is what you would EQ to achieve when each of the target headphones was measured on the target fixture, and that equalized response should sound the same as the headphone equalized to hit the Welti target (in other words the same EQ would be required to hit the Welti and the new target curve). However if the Target curves don’t look the same across headphones you know the equalized output won’t sound the same across different headphone models, the bigger the variance the larger the audible difference. (Edited to clarify)
The source for this is the interaction of the acoustic impedance of the headphone, of the 5128 ears and mics, and the 5128’s head itself. In fact the differences were quite audible, and that is why we could not produce a baseline 5128 target curve. The only headphone that would have passed the listening test with 5128 EQ would likely have been the Corina, had we tested it. The fact that the 5128 targets varied more than the KB5000 is shown by the standard deviation of the two fixtures.
The question for future investigation derives from this process. As you can see, especially for the 5128 curves, the output isn’t consistent, due both to variations in acoustic impedance of the headphones and quality of seal on the head vs flat plates. For example, the highs in the HiFiMan and 518 vary considerably more than the highs on the DCA models, and only the Corina really came out similarly on all 3. To be able to actually definitively explain that requires being able to better model the acoustic impedance of the source headphones, which is not something most audio labs are equipped to do.
Last edited:
I assume this (your?) post on Head-Fi is another view of this same project, correct?
Yes, since Amir posted a summary I didn’t want to start a thread here with the full data from the report, but anyone interested should check out that post, there is a lot more information that will help make some of the work and conclusions easier to follow.
Because I found it very informative and I thought people here might also be interested I see @Dan Clark posted an excellent follow up message in that thread on head-fi:
www.head-fi.org
Research: A New Harman Curve Target for GRAS 45CA-10 and B&K 5128
Adapting a New Harman-Equivalent Target Curve For The GRAS 45CA-10 and the B&K 5128 Since its initial release in 2013 and subsequent updates in 2018, the Harman Curve has established a general preference target curve for the headphone industry. The Harman curve should be understood to...
www.head-fi.org
Similar threads
