Love it. It clearly shows that deviation in the midrange is more noticeable (where human voices are) and due to this the preference bounds are tight. And it also shows that bass and treble really come down to personal preference, as everyone who has paid attention to how EQ works should know by now. This is a big deal, and I hope other participants in the space approach showing this data similarly. Great explanation, Andrew.
Is that right? Since an elevated treble and bass is the equivalent of a reduced midrange, couldn't you also normalise the graph around those preferences being narrow instead?
Coming back around to compensated graphs is huge. I'm so glad you've chosen to do this. The readability makes them accessible to SO many more people, and not just 'experts' or enthusiasts. Raw and compensated graphs are both flawed anyway, so why make them even more difficult to understand by leaving them raw. *[Insert Gordan Ramsey meme "ITS F****** RAW!!!"]*
Question about preference bounds. What is the distribution of "preference" within them, and how fast do they fall off at the edges? As in, if it were a gradient and not a solid shade, I doubt it would be the darkest in the very middle, and then falling off to the top and bottom edges; I guess it'll show two tendencies, like one gravitating towards "neutral" sound, and one gravitating towards "v-shaped" sound (that's probably what was mentioned briefly in the end, too) - though maybe it won't, but looking at the single-color shade, I indeed have to guess. I feel that seeing those tendencies within the bounds too would be helpful, especially for more extreme ("basshead", "treble cannon"...) deviations.
three "classes" shown at the end suggest three groups, a basshead camp, a treblehead camp, and a camp in the middle. i think Sean has a brief summary of his segmentation paper on his youtube channel. give that a look.
I appreciate this, even as a person who does not really place much value in measurements. Who knows mayhaps I'll be persuaded down the line. Thanks for the video~
The inclusion of preference bounds makes a lot of sense to me. Just out of curiosity, what percentage of listeners lie within the preference bound? 68% (one standard deviation)? 95% (two standard deviations)?
I always felt that Rin Choi's presentation of graphs was ideal (M.R.O. blog from 10 years ago). What you are doing now is very similar minus the greyed out raw data and with harman preference bounds instead of the ITU-R BS.708 tolerance mask. Overall I'm very happy with this change.
I know this video is a bit old but I am just coming across it.. As someone who spent way to much time in my theater (also listening room) doing measurements and testing for an in-room prefrence along curve (using harmon as an original baseline) this will make reading graphs and making compensation so much easier.. Although, I do feel sorry for you trying to explain this to people lol
There are a couple things I don't get. One is volume. It seems to me that somewhat different curves would be preferred at different volumes, yet volume is rarely mentioned. The other is that adjustments in say bass response might induce "new" preferences in say the upper treble, so the bass response "preference" in a target curve might be forcing preferences in another area in the curve and vice versa, so a single preference curve doesn't quite make sense to me. It's like depending on the person you might prefer more salt and pepper together, or more pepper if there's less salt, or more salt if there's less pepper, but then somehow you're trying to define the preferred combination of both, when the preferences work in tandem.
Right, so the preference research accounts for that, which is what the preference boundaries shown here also reflect. I actually think what you might be interested in is the methodology of the research itself, which is linked in the video description. Unfortunately its behind the audio engineering society paywall, but if you are interested to learn about it I highly encourage reading it. We've done some videos on it in the past and will likely continue to do more on it, but a video can only cover so much.
@@ResolveReviews Makes sense. I would like to learn more if for no other reason than to explain my own preferences. Like at medium to higher volumes, I really like the HD600, but at lower volumes, I prefer the 109 Pro. But I'm not sure I'd ultimately prefer a headphone with a response that was an average of the two. I think this speaks to your point about the target only being a reference tool and not a "goal" or the final word.
I am a bit confused. Depending on the normalization frequency point (e.g., 500 Hz or 1000 Hz), the interpretation can change a bit, especially for in-ears, right? So how do guys choose it? Averaging using white noise as seen in Harman studies, matching what is perceptually revelent to the reviewer, or another method? Another question, is the preference boundaries the same for in-ears as well? Assuming that is based on the OE research and the in-ear research suggested more bass was preferred it seems a bit unclear how the in-ear approach will be.
You're asking the right questions. Currently normalization is full range 'best fit', but we are internally debating that and potentially adding custom range normalization. For IEM data the preference boundaries will need to be wider because there's more variation in terms of IEM preferences - again based on the research. We're also going to move to a population average DFHRTF for IEMs since the pinna is being bypassed, but that work hasn't been completed yet.
You had me at Harmon curve. . Nice simplified approach to a complicated mess with a complicated approach to simplify the complications of complicated headphone measurements. . However, my approach is much simpler. Whenever a headphone measurement graph appears, I close my eyes.
I actually like what happens on sites like squiglink, where you can directly compare graphs of different headphones and you can compensate to both a target and another headphone. That imo is way more informative of similarities between different models, and is extremely useful for stuff like eq. I don't get why you don't have your own squigsite, with your database of 5128 measurements it would be an extremely useful and reliable tool for the community
I never understood preferring raw measurements. I feel like the crazy shape obscures any understanding of what the headphones even sound like. Relative to a flat line, however, the shapes become very apparent.
The question is always "how to compensate to get to a flat line" as in "what do you consider the target, from which you want to show the deviations". This is an evolving question. Harman's research is excellent but the results are not the final word (even though it likely converges very close to it, when you compare it to the outcomes of other, similarly motivated other research projects). Compensating to a strictly defined reference sound field allows for relative comparisons - and while it means that a flat line is not the goal on a DF-compensated graph, it's much easier to read still than raw (uncompensated) graphs.. That being said, I'll continue showing raw measurement results in addition to DF-compensated graphs.
@@oratory-ck1tj Yeah I'm less hardline on this than Blaine is, as I think its still fine to show raw graphs. Just that when raws are shown it should be relative to a preference target like Harman, not so much the flat DF HRTF - at that point it may as well be compensated. Like I think some people have asked us even for just the raw data, perhaps assuming erroneously that they could slide their own target or a reviewer target in there, or even compare it to other raws done on other rigs, which makes no sense for the 5128 or 4620. Those comparisons are only useful to those of us specifically interested in the differences among rigs IMO.
@@oratory-ck1tj I absolutely understand the problem of defining compensations. But even then I find that it's always an improvement to have SOME kind of compensation to make up for the huge ear gain hump. This is in the context of having seen the same headphones measured by different people on different rigs of course, as a frame of relative reference.
@@ResolveReviews Raw is fine and it still serves a purpose, but I really like that you and many other reviewers are moving towards compensated to flat DF. If you could only display one graph, the DF compensated should have priority. Review articles and videos can freely include raw measurments for those who want to dig deeper, but the graph most useful for comparison tools and least misleading for newcomers is the DF compensated plot. Preference targets are great if you can discover your own or find a reviewer that uses one you like and I think most people are ok them being included, it's just not great when graphs are made showing only raw and compensated to the preference target. If it's compensated to anything it should always be flat DF first since that's very rigurously defined and also close enough to any reasonable preference target that the sine illusion is no longer a major problem. I also think preference targets should primarily be designed and displayed on DF compensated graphs (even though this carries the risk of making it seem much more universal than it really is if the measurment rig has significant deviations from your own head in terms of the headphone transfer function). I come across a preference target displayed in raw form whithout a direct mention as to which measurement rig it was designed for far too often, and I highly doubt they are all tailor made to the HRTF of the person making the preference target (which would also make it very hard to replicate elsewhere and ultimately meaningless as a way of sharing a sound signature preference.). Either way, keep up the great work!
I really hope someone will cover the topic of lossless files, the difference between wav flac aac, and so on, and compare to 320 256 192khz. Add to that what is 24bit and 18bit when should we listen or what type of music should we listen to music based on these so many formats and so many different bitrates. Most importantly how do we spot genuine lossless files, using different software to identify the authenticity of music files could cause confusion for example Spek and Fakin' The Funk, not all the time both softwares agree with one another when accessing a lossless music files thus this caused greater confusion of am I holding a genuine flac file or not. I think someone should cover a topic like this to educate the proper basic guide maybe even before getting a headphone perhaps. Disclaimer, this message has been sent to a few audiophiles but there's no response from them yet.
Personally I rather like what you are going for with this! I think a good compromise for the people who like raw graphs is either include both in the reviews or have a link to the raw graph in the description. I agree that compensated if one understands what it is being compensated to is a good thing combined with your preference boundaries. Either way thank you for all the work you do Resolve (and of course the rest of the headphone-show team)!
ok, I will try to be calm as possible. so much angry on subjective audiophiles who don't want to understand something more complicated then potato. you right about measuring and compensation as well! and here's why.. by analogy's with measuring of studio monitors - first thing you should do is measure your room! then watch the graph with responses, try to fix them with bass traps or sound panels. measure again and again till you get the best result for your exact room. and then and only then you measure the monitors in this room and compensating it with room profile to see results. and this Harman preferable EQ is stupidity what I found recently (cos' I'm not audiophile but trying to find some exact thing for myself). it's so stupid in so much ways.. evry "bedroom producer" knows it but no one of audiophile doesn't! it's annoying how people love their hobbies but doesn't want to know something more about it. I have more to say about READING EQs, contact me if you interested. I see you really trying to research and I will help with first step.
Seems like a correct approach to me, but I am just not clear how you have compensated the preference curves / boundaries themselves. It seems to me that you would have to use the human average transfer function to adjust the Harman data, or at least something similar. Is this what was done?
Similar yes! We use the rig-specific DFHRTF, which is the ear transfer function that makes the most sense for headphones. Right now its for two 'standard' ears in the B&K 5128 and GRAS KB5000, but population average is what we'll likely go with for IEMs since the pinna is bypassed.
I don't see why we measure headphones with dummy head. I mean when we measure speakers we don't take the ear shape or ear canal into account. We assume that the brain compensates for the shape of the ear. Why don't we assume that when we measure headphones ? I'm sure that I'm missing something here but I don't know what.
It's all cool and dandy but you still don't have measurement database like like crin's. Just make a database with the squig site. That would be a massive help to community.
Where are you guys getting the data that “most” people prefer that tilted diffuse field target? Is it based on any published research? Is it basically harman “translated” to this new system? iirc we’re not using harman anymore because they did their research on an old measuring rig and we have new stuff now
The Harman research. We couldn't use the Harman Target for the new system because it was developed for the older system. But we can still use the OUTCOMES of the research, and this is what the preference boundaries show.
i just want to hear the most natural representation of instruments and the compensated graph will not show it to me. Its for masses. So whats the target for real life immersive sound?
The compensated graph is not for the masses but rather the ENTHUSIASTS. We should all prefer this view for the reasons shown in this video. The problem is, people are used to reading graphs a certain way and it will take time for them to learn.
@@Adi7x But it's not 10dB across the board.. I don't think you quite get what this is. This is literally the Harman research being put into practice. Bass preference varies by a significant amount - far more than other parts of the HRTF.
@@ResolveReviews and -10db on treble amount gives a dark muffed sound. Whos prefference is that? Thanks to such guidelines, it will be possible to justify for example, the terrible sound of the Sony WF-1000XM4, which sounds like a loudspeaker covered with a thick layer of blankets... because it fits into the defused field range. Tell me why im wrong, maby i dont understand it
@@Adi7x because this isn't compensated to a target. This is calibrated to the DFHRTF of the measurement rig. That's why it's represented as a downward tilt. So in other words, a flat line is NOT what anyone wants here because that's the DFHRTF. The preference boundaries are pulled directly from the Harman research.
That all makes sense however it would also be interesting to see the compensation graph as a dotted line or something. It wouldn't have to be a dotted line. It could be a line made up of tiny little swastikas if you prefer.
I feel this is a very good direction this video takes. This is both better than Harmon curve, plus it matches the range of signatures that the most expensive headphones fall in (that I tried). I myself have done countless measuring and altering many headphones patameters so like to point out an important aspect. IMO, from my own experience in testing many Top headphones in past, I found that most headphones measurements helped with signature preference. Like this good video tooic, they are goodbat telling the characters and signature, but actually not the performance. Very few measurements actually correlated with performance and most importantly "realism". In fact there were specific "distortions" inherent to divers that were pleasing, such as the sound of a "planar". Regardless, what I wanted to point out, was that the best headphones all had only ONE measurement, which was similarly superior to all other headphones. That specific trait was a " large" sound, combined with SPEED... The impulse response didn't even have to be "perfect", as distortions or deviations in time actually disclosed its driver signature, not actual "imperfections". So this very important aspect of speed, is why they had realism, and is also why we see many IEMs being high in realism, yet I found very few of the best IEMs to come close to matching actual headphones in sheer size/depth/spaciousness of image. Most are too forward & deep, unlike the more balanced imaging of headphones.
Please forgive me if this is a stupid question. This RUclipsr: www.youtube.com/@askdrtk/videos - uses sine sweeps to evaluate the frequency response of headphones. Wouldn't that be raw data? Wouldn't raw frequency response information be a better place to start a comparative analysis? Thanks.
Ok, I didn't listen far enough, but you didn't explain it well, or perhaps I should say, you didn't explain it at all. You know, this lack of explanation makes a great deal of the information you've delineated on your channel a bit suspect. Of course, you've neglected spoken voice, so there is that too of course.
This is sound in a reverberation chamber, where the sound pressure is equal throughout the room, and where the reflected sound dominates the direct sound. It is the opposite of a free field, which is sound in an anechoic chamber.
I've been looking at that graphical perception chart for a few minutes and I don't get it. Anyone cares to explain and the relation it has to FR graphs.
When reading a FR graph, the main thing we care about is the difference in SPL (dB) between the measurement and the target, or in other words the vertical distance between the two on the graph. Now for example, look at the top left graph in the graphical perception chart. Imagine you have a ruler positioned perfectly vertically from top to bottom, and slide it from left to right across the graph. You would find that although it appears that the lines converge, they actually have the same vertical distance from one another at all points across the x-axis, or in the case of a FR graph, each frequency. So when compensated, the top left graph corresponds to the 4th graph from the left in the top row, the flat line. So what Resolve is saying is that compensated FR graphs are easier for us to read accurately compared to raw graphs, particularly in the ear-gain region, because of this error in our perception. Hope this helps!
@@nshr1694 Thanks for the reply! That makes sense, at least when comparing the other figures! Most of the sites comparing graphs have an option to show the difference in dB compared to target (and/or other devices) when hovering the cursor over the lines, but I guess that still makes it hard to tell where the actual changes are? I've noticed something similar when looking at Crinacle's normalizing between 60dB (for example) and 1000Hz. Do you know how that works?
@@Ca11mero No problem! I believe you are referring to Crinacle's article "Graphs 101: How to Read Headphone Measurements" ? I'm no data scientist, but it's true that the way a graph is normalized (at what point the two graphs are equal) also has an effect on how measurements are perceived. Similar to how calling a headphone "dark" (less highs) or calling it "warm" (more lower mid/bass) are nearly the same thing. (And even then these words might not mean exactly the same thing to each person) As Crinacle said in his article, normalizing at 1000Hz is more of a rule of thumb and there is no one rule when it comes to normalization. Normalizing the graphs by perceived volume (equal loudness) is another method Crinacle mentioned and this would work better for some cases as well. In the end it's up to the reviewer or source of the measurement to make sure the data is presented in a way that makes sense. Since FR measurements are just a curve we are really not comparing the differences in the values at specific points like I said earlier, but instead the shape of the two curves overall. For that reason there is no one single way to place a measured frequency response and a preference curve on the same graph for comparison. Again I am no expert on data presentation, but I hope I was able to help you understand! And always remember, you do not have the same ears as the test rig and the same preferences as the reviewer, so it's always best to listen in-person!
@@nshr1694 Thanks a lot again, it's very interesting. Yes, that was what I was thinking about! But I'm not sure why you would normalize at a certain Hz rather than volume. Interesting, so what's why you should stay away from stuff like AutoEQ? Since it just matches the curve? So there is no real scientific way to EQ to a target I assume. Also there doesn't seem to be any good way to figure out where peaks are besides guessing it's somewhere around where it's been measured. Some people suggest tone generators (listening to specific Hz) or sweeps but that has it's own set of problems. Like a guitar note playing is a set of frequencies rather than a single frequency. Yes, that I know of! :)
@@Ca11mero Haha to be honest you've begun to get me over my head here. Crinacle does a good job explaining why normalizing at a certain frequency might make more sense in a few cases rather then normalizing by volume when he compares the ThieAudio Voyager 3 settings. But personally I think it would be fine to normalize by volume as the differences could also be heard as a boost in the lows rather than a cut in the highs. But what Crinacle does when he normalizes by 200hz in his last example is make the data easier to read. For AutoEQ, I think it is a fine tool. You just need to remember that targets such as the Harman curve are just what most people are likely to prefer, and it is probably not going to match your preferences perfectly, which is part of what Resolve was talking about in this video with the new preference bounds in these graphs. I think AutoEQ is a good baseline to fix parts of the FR that are "incorrect" but at the very least one should also be adjusting bass and treble levels to preference. I've personally experimented with EQing my dt880 600ohm and I've found that EQ's such as oratory1990's and Amir's from ASR simply do not sound great to me. Whether this is because of unit variation in the headphone, or because of my ears or my signal chain I don't know, but I do know that the treble in these EQ's does not sound correct to my ears. I've found that tone generators can help in identifying major peaks or dips, but I think in the end it is most important for whatever kind of equalization you are doing to sound good in the music you are listening to.
The preference boundaries are definitely a nice addition, but the removal of the df+slope line from the graph feels like a big step backwards in readability. I'm excited about the preference boundaries because too many people in this hobby view measurements purely from the perspective of "deviation from target=bad", but by removing the df+slope line, a ton of ambiguity is created in where these preferences exist relative to the dfhrtf of this particular head. At least for me, the graph shown at 3:08 featuring both the df+slope and where the preference bounds exist in relation to it is by far the most readable and gives me the best reference for the kind of sound that the preference bounds represent. tldr: I think the preference bounds are great, but removing other reference points makes them ambiguous and makes the measurements harder to read, at least for me
Now we need to get to the bottom of the headphone cable controversy. WaveTheory has been flat out making videos on various headphone cables. Do some people belong in the looney bin or could it be that the force is real and headphone cables make tiny differences in the force.
Nice work to you all! Only thing I would say is that for hifiman headphones and other similar headphones the preference boundaries for the pinna gain seems to only partially apply since it is shifted . I think must people would not touch the pinna gain of the hifiman, but would touch up the trebble and bass. I would maybe add another shaded area for shifted pinna gain headphones.
@@Ca11mero with the sundara it is a bit less present, bit with the arya etc the pinna gain is shifted from 1k to 2k. Which is in part responsible of the enlarged soundstage impression
I'd love to see you guys put something in a measurement tool like squig so people can still do what they want and are used to though. Truth is especially for headphones where the research is stronger, having what I assume ended up being the target (average, or standard deviation?) preferred by the majority within those bounds is still useful. I realise the 5128 is forcing your hands here... End of the day I'll continue just going elsewhere for measurements though for headphones because it's what I'm used to. I know what Harman sounds like. Still interesting to see if your rather heavy handed approach to usage of 5128 comes out with something new.
Even within the largest group there were still preference 'clusters'. I do know what you mean about this, and we're still considering different shading or maybe even a gradient to reflect that.
Isn't the Harman target only for the older rig and thus by definition missing out on the benefits of the 5128? Seemed to me it's worthwhile to get used to the new graphs even when I myself am starting to measure with a clone 711.
@@hartyewh1 Yeah I think we're just at a super awkward spot where the only actual research is on 711, therefore it still provides more overall benefits.
Now all we need is for subjectivists to comment on how they hear the thing relative to the compensated frequency response and also convince people who are scared of EQ that it's not the work of the Devil.
Your speech is too quick. Hence it is difficult to follow you, especially for someone who is not familiar with the topic. You also make quite much mistakes in your expressions. Hence it would be good if you paid more attention to the correctness of your expressions, both for you and for the public terminology and also that would decrease your speed, which would be positive IMO. You do not "compensate measurement", you compensate graph of frequency response. Not '"measurements' but 'frequency responses' are represented. Frequency response is measured, so measurement's output is frequency response. They are not the same concepts, one is the subject or output of the other. They are not "represented" differently but "presented" differently. Their representation is quite uniform: an array of real numbers [probably floating point, in computers]. The 'target' is a frequency response, not a curve. The target, as FR, can be and often is presented by a curve, thus the "target curve" expression is valid, but that does not make the target a curve.
If someone's speech is too quick for you, then RUclips has some neat accessibility features to make it easier to comprehend. You can slow the video down, or turn on auto generated subtitles which are surprisingly accurate. Everyone's speech has a different cadence and flow to it, it's not really something you can control over extended periods of time while still retaining normal mental faculties. And it isn't inaccurate to say that a measurement is compensated. The measurement of the frequency response is what we are directly seeing. The actual frequency response is not being altered, simply how it is being represented in the measurements. Even if saying that were a mistake, mostly everyone understands what the intent is without too much trouble. In the same way that if someone says "you make quite much mistakes" instead of "you make quite many mistakes," people just instinctively understand what the actual intended meaning is. Because making tiny errors in our communication is just part of being human.
@@millsjonah We should weigh mistakes by their severity, severity is not countable. 2 mistake of the same type - is it 1 mistake or 2 mistakes? I would say somewhere in the middle. ...
This is great! Preference boundaries was something we definitely needed. This will make it much easier for people to grasp the data
Boundaries makes SO much more sense than 'target curves' which never made sense to me.
@@GREQReviews Couldn't agree more.
Love it. It clearly shows that deviation in the midrange is more noticeable (where human voices are) and due to this the preference bounds are tight. And it also shows that bass and treble really come down to personal preference, as everyone who has paid attention to how EQ works should know by now.
This is a big deal, and I hope other participants in the space approach showing this data similarly.
Great explanation, Andrew.
Is that right? Since an elevated treble and bass is the equivalent of a reduced midrange, couldn't you also normalise the graph around those preferences being narrow instead?
This is actually quiet a straight forward and understandable explanation of what you've done.
Coming back around to compensated graphs is huge.
I'm so glad you've chosen to do this.
The readability makes them accessible to SO many more people, and not just 'experts' or enthusiasts.
Raw and compensated graphs are both flawed anyway, so why make them even more difficult to understand by leaving them raw.
*[Insert Gordan Ramsey meme "ITS F****** RAW!!!"]*
I came to this videos almost mad that you guys are changing to compensated, but you actually convinced me. Love you videos
I appreciate you giving this a chance!
Preference boundaries are a great addition. Measurements being taken out of context is a big problem within the community.
Question about preference bounds. What is the distribution of "preference" within them, and how fast do they fall off at the edges? As in, if it were a gradient and not a solid shade, I doubt it would be the darkest in the very middle, and then falling off to the top and bottom edges; I guess it'll show two tendencies, like one gravitating towards "neutral" sound, and one gravitating towards "v-shaped" sound (that's probably what was mentioned briefly in the end, too) - though maybe it won't, but looking at the single-color shade, I indeed have to guess. I feel that seeing those tendencies within the bounds too would be helpful, especially for more extreme ("basshead", "treble cannon"...) deviations.
three "classes" shown at the end suggest three groups, a basshead camp, a treblehead camp, and a camp in the middle. i think Sean has a brief summary of his segmentation paper on his youtube channel. give that a look.
I appreciate this, even as a person who does not really place much value in measurements. Who knows mayhaps I'll be persuaded down the line. Thanks for the video~
The inclusion of preference bounds makes a lot of sense to me. Just out of curiosity, what percentage of listeners lie within the preference bound? 68% (one standard deviation)? 95% (two standard deviations)?
References are in the description
luckily i've listened to you and crew nerd out about this stuff enough that I understand what is explained here
I always felt that Rin Choi's presentation of graphs was ideal (M.R.O. blog from 10 years ago). What you are doing now is very similar minus the greyed out raw data and with harman preference bounds instead of the ITU-R BS.708 tolerance mask.
Overall I'm very happy with this change.
I know this video is a bit old but I am just coming across it.. As someone who spent way to much time in my theater (also listening room) doing measurements and testing for an in-room prefrence along curve (using harmon as an original baseline) this will make reading graphs and making compensation so much easier.. Although, I do feel sorry for you trying to explain this to people lol
There are a couple things I don't get. One is volume. It seems to me that somewhat different curves would be preferred at different volumes, yet volume is rarely mentioned. The other is that adjustments in say bass response might induce "new" preferences in say the upper treble, so the bass response "preference" in a target curve might be forcing preferences in another area in the curve and vice versa, so a single preference curve doesn't quite make sense to me. It's like depending on the person you might prefer more salt and pepper together, or more pepper if there's less salt, or more salt if there's less pepper, but then somehow you're trying to define the preferred combination of both, when the preferences work in tandem.
Right, so the preference research accounts for that, which is what the preference boundaries shown here also reflect. I actually think what you might be interested in is the methodology of the research itself, which is linked in the video description. Unfortunately its behind the audio engineering society paywall, but if you are interested to learn about it I highly encourage reading it. We've done some videos on it in the past and will likely continue to do more on it, but a video can only cover so much.
@@ResolveReviews Makes sense. I would like to learn more if for no other reason than to explain my own preferences. Like at medium to higher volumes, I really like the HD600, but at lower volumes, I prefer the 109 Pro. But I'm not sure I'd ultimately prefer a headphone with a response that was an average of the two. I think this speaks to your point about the target only being a reference tool and not a "goal" or the final word.
Sennheiser explains how the tuning of the ie900 is meant to be suited for low volume listening
Very nice!
Love this. Leave it to you guys to finally figure it out
any way to measure these spatial enhancement, ala Apple, Bose, Sony, maybe on some other audiophile wireless offerings?
I am a bit confused. Depending on the normalization frequency point (e.g., 500 Hz or 1000 Hz), the interpretation can change a bit, especially for in-ears, right? So how do guys choose it? Averaging using white noise as seen in Harman studies, matching what is perceptually revelent to the reviewer, or another method?
Another question, is the preference boundaries the same for in-ears as well? Assuming that is based on the OE research and the in-ear research suggested more bass was preferred it seems a bit unclear how the in-ear approach will be.
You're asking the right questions. Currently normalization is full range 'best fit', but we are internally debating that and potentially adding custom range normalization. For IEM data the preference boundaries will need to be wider because there's more variation in terms of IEM preferences - again based on the research. We're also going to move to a population average DFHRTF for IEMs since the pinna is being bypassed, but that work hasn't been completed yet.
You had me at Harmon curve.
.
Nice simplified approach to a complicated mess with a complicated approach to simplify the complications of complicated headphone measurements.
.
However, my approach is much simpler. Whenever a headphone measurement graph appears, I close my eyes.
With this new method of displaying measurements, you may find graphs to be more valuable. I encourage at least giving it a good faith try.
I actually like what happens on sites like squiglink, where you can directly compare graphs of different headphones and you can compensate to both a target and another headphone. That imo is way more informative of similarities between different models, and is extremely useful for stuff like eq. I don't get why you don't have your own squigsite, with your database of 5128 measurements it would be an extremely useful and reliable tool for the community
I never understood preferring raw measurements. I feel like the crazy shape obscures any understanding of what the headphones even sound like. Relative to a flat line, however, the shapes become very apparent.
The question is always "how to compensate to get to a flat line" as in "what do you consider the target, from which you want to show the deviations".
This is an evolving question. Harman's research is excellent but the results are not the final word (even though it likely converges very close to it, when you compare it to the outcomes of other, similarly motivated other research projects).
Compensating to a strictly defined reference sound field allows for relative comparisons - and while it means that a flat line is not the goal on a DF-compensated graph, it's much easier to read still than raw (uncompensated) graphs..
That being said, I'll continue showing raw measurement results in addition to DF-compensated graphs.
@@oratory-ck1tj Yeah I'm less hardline on this than Blaine is, as I think its still fine to show raw graphs. Just that when raws are shown it should be relative to a preference target like Harman, not so much the flat DF HRTF - at that point it may as well be compensated. Like I think some people have asked us even for just the raw data, perhaps assuming erroneously that they could slide their own target or a reviewer target in there, or even compare it to other raws done on other rigs, which makes no sense for the 5128 or 4620. Those comparisons are only useful to those of us specifically interested in the differences among rigs IMO.
@@oratory-ck1tj I absolutely understand the problem of defining compensations. But even then I find that it's always an improvement to have SOME kind of compensation to make up for the huge ear gain hump. This is in the context of having seen the same headphones measured by different people on different rigs of course, as a frame of relative reference.
@@ResolveReviews Raw is fine and it still serves a purpose, but I really like that you and many other reviewers are moving towards compensated to flat DF. If you could only display one graph, the DF compensated should have priority. Review articles and videos can freely include raw measurments for those who want to dig deeper, but the graph most useful for comparison tools and least misleading for newcomers is the DF compensated plot.
Preference targets are great if you can discover your own or find a reviewer that uses one you like and I think most people are ok them being included, it's just not great when graphs are made showing only raw and compensated to the preference target. If it's compensated to anything it should always be flat DF first since that's very rigurously defined and also close enough to any reasonable preference target that the sine illusion is no longer a major problem. I also think preference targets should primarily be designed and displayed on DF compensated graphs (even though this carries the risk of making it seem much more universal than it really is if the measurment rig has significant deviations from your own head in terms of the headphone transfer function). I come across a preference target displayed in raw form whithout a direct mention as to which measurement rig it was designed for far too often, and I highly doubt they are all tailor made to the HRTF of the person making the preference target (which would also make it very hard to replicate elsewhere and ultimately meaningless as a way of sharing a sound signature preference.). Either way, keep up the great work!
I really hope someone will cover the topic of lossless files, the difference between wav flac aac, and so on, and compare to 320 256 192khz. Add to that what is 24bit and 18bit when should we listen or what type of music should we listen to music based on these so many formats and so many different bitrates. Most importantly how do we spot genuine lossless files, using different software to identify the authenticity of music files could cause confusion for example Spek and Fakin' The Funk, not all the time both softwares agree with one another when accessing a lossless music files thus this caused greater confusion of am I holding a genuine flac file or not. I think someone should cover a topic like this to educate the proper basic guide maybe even before getting a headphone perhaps.
Disclaimer, this message has been sent to a few audiophiles but there's no response from them yet.
Personally I rather like what you are going for with this!
I think a good compromise for the people who like raw graphs is either include both in the reviews or have a link to the raw graph in the description.
I agree that compensated if one understands what it is being compensated to is a good thing combined with your preference boundaries.
Either way thank you for all the work you do Resolve (and of course the rest of the headphone-show team)!
Great video! So when we see a "raw measurement vs Harman target curve" it's actually "raw vs average of Harman target curve"?
Yes basically
ok, I will try to be calm as possible. so much angry on subjective audiophiles who don't want to understand something more complicated then potato.
you right about measuring and compensation as well! and here's why..
by analogy's with measuring of studio monitors - first thing you should do is measure your room! then watch the graph with responses, try to fix them with bass traps or sound panels. measure again and again till you get the best result for your exact room. and then and only then you measure the monitors in this room and compensating it with room profile to see results.
and this Harman preferable EQ is stupidity what I found recently (cos' I'm not audiophile but trying to find some exact thing for myself). it's so stupid in so much ways..
evry "bedroom producer" knows it but no one of audiophile doesn't! it's annoying how people love their hobbies but doesn't want to know something more about it.
I have more to say about READING EQs, contact me if you interested. I see you really trying to research and I will help with first step.
Seems like a correct approach to me, but I am just not clear how you have compensated the preference curves / boundaries themselves. It seems to me that you would have to use the human average transfer function to adjust the Harman data, or at least something similar. Is this what was done?
Similar yes! We use the rig-specific DFHRTF, which is the ear transfer function that makes the most sense for headphones. Right now its for two 'standard' ears in the B&K 5128 and GRAS KB5000, but population average is what we'll likely go with for IEMs since the pinna is bypassed.
I appreciate this
I don't see why we measure headphones with dummy head. I mean when we measure speakers we don't take the ear shape or ear canal into account. We assume that the brain compensates for the shape of the ear. Why don't we assume that when we measure headphones ? I'm sure that I'm missing something here but I don't know what.
Nice 👍
Thanks really helpful and very useful to evaluate headphones.
It's all cool and dandy but you still don't have measurement database like like crin's. Just make a database with the squig site. That would be a massive help to community.
Finally! Thank you very much for this change, it makes a lot more sense to me 😊
Where are you guys getting the data that “most” people prefer that tilted diffuse field target? Is it based on any published research? Is it basically harman “translated” to this new system? iirc we’re not using harman anymore because they did their research on an old measuring rig and we have new stuff now
The Harman research. We couldn't use the Harman Target for the new system because it was developed for the older system. But we can still use the OUTCOMES of the research, and this is what the preference boundaries show.
@@ResolveReviews ahhhh, that makes sense! Fascinating! Do you have a video explaining the advantage of the new system over the old one?
@@aGrimTurtle Well, that's the intention of this video. But we will need to do more.
@@ResolveReviews Gotcha, I'll give it a re-watch with this info in mind. Thanks for taking the time to answer questions!
i just want to hear the most natural representation of instruments and the compensated graph will not show it to me. Its for masses. So whats the target for real life immersive sound?
The compensated graph is not for the masses but rather the ENTHUSIASTS. We should all prefer this view for the reasons shown in this video. The problem is, people are used to reading graphs a certain way and it will take time for them to learn.
@@TheHEADPHONEShow i think the new target is more forgiving and reviewers can fit more headphones into the graph. +-10 db is a huge tolerance
@@Adi7x But it's not 10dB across the board.. I don't think you quite get what this is. This is literally the Harman research being put into practice. Bass preference varies by a significant amount - far more than other parts of the HRTF.
@@ResolveReviews and -10db on treble amount gives a dark muffed sound. Whos prefference is that? Thanks to such guidelines, it will be possible to justify for example, the terrible sound of the Sony WF-1000XM4, which sounds like a loudspeaker covered with a thick layer of blankets... because it fits into the defused field range. Tell me why im wrong, maby i dont understand it
@@Adi7x because this isn't compensated to a target. This is calibrated to the DFHRTF of the measurement rig. That's why it's represented as a downward tilt. So in other words, a flat line is NOT what anyone wants here because that's the DFHRTF. The preference boundaries are pulled directly from the Harman research.
Yes.
how will compensation be done?
Can/will y'all make a (very niche) video on how to capture DF comp for those of us who want to adhere to this standard?
That all makes sense however it would also be interesting to see the compensation graph as a dotted line or something. It wouldn't have to be a dotted line. It could be a line made up of tiny little swastikas if you prefer.
I feel this is a very good direction this video takes. This is both better than Harmon curve, plus it matches the range of signatures that the most expensive headphones fall in (that I tried).
I myself have done countless measuring and altering many headphones patameters so like to point out an important aspect.
IMO, from my own experience in testing many Top headphones in past, I found that most headphones measurements helped with signature preference. Like this good video tooic, they are goodbat telling the characters and signature, but actually not the performance.
Very few measurements actually correlated with performance and most importantly "realism". In fact there were specific "distortions" inherent to divers that were pleasing, such as the sound of a "planar".
Regardless, what I wanted to point out, was that the best headphones all had only ONE measurement, which was similarly superior to all other headphones.
That specific trait was a " large" sound, combined with SPEED...
The impulse response didn't even have to be "perfect", as distortions or deviations in time actually disclosed its driver signature, not actual "imperfections".
So this very important aspect of speed, is why they had realism, and is also why we see many IEMs being high in realism, yet I found very few of the best IEMs to come close to matching actual headphones in sheer size/depth/spaciousness of image.
Most are too forward & deep, unlike the more balanced imaging of headphones.
seems good to me
Please forgive me if this is a stupid question. This RUclipsr: www.youtube.com/@askdrtk/videos - uses sine sweeps to evaluate the frequency response of headphones. Wouldn't that be raw data? Wouldn't raw frequency response information be a better place to start a comparative analysis? Thanks.
Ok, I didn't listen far enough, but you didn't explain it well, or perhaps I should say, you didn't explain it at all. You know, this lack of explanation makes a great deal of the information you've delineated on your channel a bit suspect. Of course, you've neglected spoken voice, so there is that too of course.
I'm listening to you speaking with headphones actually, You know, spoken voice.
Hooray!!!!! As a data scientist I really appreciate this. Great work 🙂
I like this much better.
What does "diffuse field" mean?
This is sound in a reverberation chamber, where the sound pressure is equal throughout the room, and where the reflected sound dominates the direct sound. It is the opposite of a free field, which is sound in an anechoic chamber.
I've been looking at that graphical perception chart for a few minutes and I don't get it.
Anyone cares to explain and the relation it has to FR graphs.
When reading a FR graph, the main thing we care about is the difference in SPL (dB) between the measurement and the target, or in other words the vertical distance between the two on the graph.
Now for example, look at the top left graph in the graphical perception chart. Imagine you have a ruler positioned perfectly vertically from top to bottom, and slide it from left to right across the graph. You would find that although it appears that the lines converge, they actually have the same vertical distance from one another at all points across the x-axis, or in the case of a FR graph, each frequency. So when compensated, the top left graph corresponds to the 4th graph from the left in the top row, the flat line.
So what Resolve is saying is that compensated FR graphs are easier for us to read accurately compared to raw graphs, particularly in the ear-gain region, because of this error in our perception.
Hope this helps!
@@nshr1694 Thanks for the reply!
That makes sense, at least when comparing the other figures! Most of the sites comparing graphs have an option to show the difference in dB compared to target (and/or other devices) when hovering the cursor over the lines, but I guess that still makes it hard to tell where the actual changes are?
I've noticed something similar when looking at Crinacle's normalizing between 60dB (for example) and 1000Hz. Do you know how that works?
@@Ca11mero No problem!
I believe you are referring to Crinacle's article "Graphs 101: How to Read Headphone Measurements" ?
I'm no data scientist, but it's true that the way a graph is normalized (at what point the two graphs are equal) also has an effect on how measurements are perceived. Similar to how calling a headphone "dark" (less highs) or calling it "warm" (more lower mid/bass) are nearly the same thing. (And even then these words might not mean exactly the same thing to each person)
As Crinacle said in his article, normalizing at 1000Hz is more of a rule of thumb and there is no one rule when it comes to normalization. Normalizing the graphs by perceived volume (equal loudness) is another method Crinacle mentioned and this would work better for some cases as well.
In the end it's up to the reviewer or source of the measurement to make sure the data is presented in a way that makes sense. Since FR measurements are just a curve we are really not comparing the differences in the values at specific points like I said earlier, but instead the shape of the two curves overall. For that reason there is no one single way to place a measured frequency response and a preference curve on the same graph for comparison.
Again I am no expert on data presentation, but I hope I was able to help you understand! And always remember, you do not have the same ears as the test rig and the same preferences as the reviewer, so it's always best to listen in-person!
@@nshr1694 Thanks a lot again, it's very interesting.
Yes, that was what I was thinking about! But I'm not sure why you would normalize at a certain Hz rather than volume.
Interesting, so what's why you should stay away from stuff like AutoEQ? Since it just matches the curve? So there is no real scientific way to EQ to a target I assume. Also there doesn't seem to be any good way to figure out where peaks are besides guessing it's somewhere around where it's been measured. Some people suggest tone generators (listening to specific Hz) or sweeps but that has it's own set of problems. Like a guitar note playing is a set of frequencies rather than a single frequency.
Yes, that I know of! :)
@@Ca11mero Haha to be honest you've begun to get me over my head here. Crinacle does a good job explaining why normalizing at a certain frequency might make more sense in a few cases rather then normalizing by volume when he compares the ThieAudio Voyager 3 settings. But personally I think it would be fine to normalize by volume as the differences could also be heard as a boost in the lows rather than a cut in the highs. But what Crinacle does when he normalizes by 200hz in his last example is make the data easier to read.
For AutoEQ, I think it is a fine tool. You just need to remember that targets such as the Harman curve are just what most people are likely to prefer, and it is probably not going to match your preferences perfectly, which is part of what Resolve was talking about in this video with the new preference bounds in these graphs. I think AutoEQ is a good baseline to fix parts of the FR that are "incorrect" but at the very least one should also be adjusting bass and treble levels to preference.
I've personally experimented with EQing my dt880 600ohm and I've found that EQ's such as oratory1990's and Amir's from ASR simply do not sound great to me. Whether this is because of unit variation in the headphone, or because of my ears or my signal chain I don't know, but I do know that the treble in these EQ's does not sound correct to my ears. I've found that tone generators can help in identifying major peaks or dips, but I think in the end it is most important for whatever kind of equalization you are doing to sound good in the music you are listening to.
The preference boundaries are definitely a nice addition, but the removal of the df+slope line from the graph feels like a big step backwards in readability. I'm excited about the preference boundaries because too many people in this hobby view measurements purely from the perspective of "deviation from target=bad", but by removing the df+slope line, a ton of ambiguity is created in where these preferences exist relative to the dfhrtf of this particular head. At least for me, the graph shown at 3:08 featuring both the df+slope and where the preference bounds exist in relation to it is by far the most readable and gives me the best reference for the kind of sound that the preference bounds represent. tldr: I think the preference bounds are great, but removing other reference points makes them ambiguous and makes the measurements harder to read, at least for me
We may still add the 10dB line in there (or use shading or gradients to accomplish something similar).
Now we need to get to the bottom of the headphone cable controversy. WaveTheory has been flat out making videos on various headphone cables. Do some people belong in the looney bin or could it be that the force is real and headphone cables make tiny differences in the force.
Cables make people feel special....
@@chefsteve8381 do you feel special with special cables?
@@dangerzone007 i am special with and without cables...
@@chefsteve8381 of course you are.
Nice work to you all! Only thing I would say is that for hifiman headphones and other similar headphones the preference boundaries for the pinna gain seems to only partially apply since it is shifted . I think must people would not touch the pinna gain of the hifiman, but would touch up the trebble and bass. I would maybe add another shaded area for shifted pinna gain headphones.
What do you mean by "shifted" pinna gain?
@@Ca11mero with the sundara it is a bit less present, bit with the arya etc the pinna gain is shifted from 1k to 2k. Which is in part responsible of the enlarged soundstage impression
you need to make a more in depth video, at least 40 minutes.
I like this new line, because I have different preferences within a certain range rather than an exact line.
I'd love to see you guys put something in a measurement tool like squig so people can still do what they want and are used to though.
Truth is especially for headphones where the research is stronger, having what I assume ended up being the target (average, or standard deviation?) preferred by the majority within those bounds is still useful. I realise the 5128 is forcing your hands here... End of the day I'll continue just going elsewhere for measurements though for headphones because it's what I'm used to. I know what Harman sounds like.
Still interesting to see if your rather heavy handed approach to usage of 5128 comes out with something new.
Even within the largest group there were still preference 'clusters'. I do know what you mean about this, and we're still considering different shading or maybe even a gradient to reflect that.
Isn't the Harman target only for the older rig and thus by definition missing out on the benefits of the 5128? Seemed to me it's worthwhile to get used to the new graphs even when I myself am starting to measure with a clone 711.
@@hartyewh1 Yeah I think we're just at a super awkward spot where the only actual research is on 711, therefore it still provides more overall benefits.
@@lazar1an True, it's for the future for sure.
Give it to me raw 😋😝
Now all we need is for subjectivists to comment on how they hear the thing relative to the compensated frequency response and also convince people who are scared of EQ that it's not the work of the Devil.
Your speech is too quick. Hence it is difficult to follow you, especially for someone who is not familiar with the topic. You also make quite much mistakes in your expressions. Hence it would be good if you paid more attention to the correctness of your expressions, both for you and for the public terminology and also that would decrease your speed, which would be positive IMO.
You do not "compensate measurement", you compensate graph of frequency response.
Not '"measurements' but 'frequency responses' are represented. Frequency response is measured, so measurement's output is frequency response. They are not the same concepts, one is the subject or output of the other.
They are not "represented" differently but "presented" differently. Their representation is quite uniform: an array of real numbers [probably floating point, in computers].
The 'target' is a frequency response, not a curve. The target, as FR, can be and often is presented by a curve, thus the "target curve" expression is valid, but that does not make the target a curve.
If someone's speech is too quick for you, then RUclips has some neat accessibility features to make it easier to comprehend. You can slow the video down, or turn on auto generated subtitles which are surprisingly accurate. Everyone's speech has a different cadence and flow to it, it's not really something you can control over extended periods of time while still retaining normal mental faculties. And it isn't inaccurate to say that a measurement is compensated. The measurement of the frequency response is what we are directly seeing. The actual frequency response is not being altered, simply how it is being represented in the measurements. Even if saying that were a mistake, mostly everyone understands what the intent is without too much trouble. In the same way that if someone says "you make quite much mistakes" instead of "you make quite many mistakes," people just instinctively understand what the actual intended meaning is. Because making tiny errors in our communication is just part of being human.
@@millsjonah
"Many" is for countable quantities. The mistakes that i was talking about are not countable.
@@zoltantoth1566 yes, they are. Mistakes are a countable entity. You can make 1 mistake, you can make 2 mistakes. You're just incorrect
@@millsjonah
We should weigh mistakes by their severity, severity is not countable.
2 mistake of the same type - is it 1 mistake or 2 mistakes? I would say somewhere in the middle.
...
@@zoltantoth1566 just admit you made a mistake dude, it doesn't matter in the slightest.