[jaakkopasanen]

I've been working on equalizing headphones based on frequency responses. So far I am compensating the response for harman target, inverting the response, limiting maximum gain and smoothing out sharp edges introduced by gain clipping. What should I take into consideration when producing the equalization curve to make it sound as good as possible? I've read somewhere that it might not be best to smooth out all the small and narrow pits and bumps but I have yet to hear a convincing reason or evidence for that. Is it not a headphone equalized to harman target as good as that model can get or is there something else in play, what?

 

[bigshot]

The way I EQ is to work from a normal listening volume up to a loud one in passes. That way I don't rip my ears up. Big spikes are the most important to fix. Next in priority is the overall curve. If you have bands left over in your parametric equalizer, then you can work on the little bumps.

Abstract charts like the Harman curve are a great starting point. But you shouldn't be afraid to make adjustments to that to suit your cans and ears. The best way to train yourself to EQ is to work on identifying what bands sound like in music. That can take a while, but once you have that down, you can do critical listening to a reference track and know pretty accurately what you should fix.

 

[jaakkopasanen]

Thanks for the reply. Those are good advises I'm sure. Unfortunately I don't have the luxury of listening to headphones I'm equalizing because there are over 1000 models I'm working on.This means the equalization curve creation has to be purely algorithmic. Also I'm making these at the moment for Equalizer APO's graphic eq which is based on convolution and can work with arbitrary shaped curves and number of points. I suppose I could add an optimizer to adjust parametric eq filters or even like 10 band fixed eq. This is the reason I'm asking for theoretical foundation on building the curves for optimal sound quality. Equalizing to Harman target is a simple approach but I'm wondering if it's the best.

I got most things sorted out but still need to work a bit on the compensation. I will post sample results for some headphones shortly, and dump them all (I guess 1100 - 1200) when I'm done.

 

[castleofargh]

if you make impulses for the convolver of EQ APO, then you don't have to worry about the number of points you can afford to EQ. do you measure the headphones yourself or you rely on online graph? if so how confident are you that you're really applying the Harman target compensation?

the main issue with very precise EQ from headphone measurements comes from the variations from production. if the difference ends up being a few dB above or below somewhere, no big deal. but if the variations is something like a little shift in frequency for a resonance peak, then an EQ could end up making things worst on several pairs of that model. not dealing with the resonance properly, and attenuating the signal right next to it.

I would also go easy on changes above say 8 or 9khz, simply for how much variations can exist based on how we put the headphone on, and how different our own head can be from the measurement rig. so same as above, but worst, I'd avoid trying to fix massive and rather local issues, or only apply a moderate correction just in case.

 

[bigshot]

Yeah. Any canned EQ setting is going to be a spit in the wind. It's still probably going to need to be tweaked.

 

[jaakkopasanen]

if you make impulses for the convolver of EQ APO, then you don't have to worry about the number of points you can afford to EQ. do you measure the headphones yourself or you rely on online graph? if so how confident are you that you're really applying the Harman target compensation?

the main issue with very precise EQ from headphone measurements comes from the variations from production. if the difference ends up being a few dB above or below somewhere, no big deal. but if the variations is something like a little shift in frequency for a resonance peak, then an EQ could end up making things worst on several pairs of that model. not dealing with the resonance properly, and attenuating the signal right next to it.

I would also go easy on changes above say 8 or 9khz, simply for how much variations can exist based on how we put the headphone on, and how different our own head can be from the measurement rig. so same as above, but worst, I'd avoid trying to fix massive and rather local issues, or only apply a moderate correction just in case.

Equalizer APO's graphic eq does the impulse responses on the fly, so that's convenient. I'm using measurements from headphone.com and Innerfidelity. Problem with headphone.com measurements is that they have not calibrated their system for Harman target. Fortunately Tyll of Innerfidelity has done so and already published his new compensation curve. I'm thinking of getting all the same models measured by both and finding an average compensation from headphone.com's measurements to Innerfidelity's Harman target compensated curves.

That resonance shift in frequency axis is definitely something to worry about. I remember Tyll mentioning that the spikes in 10 kHz region are caused by ear canal resonances and shift in the frequency axis as headphone position changes. I have Savitzky-Golay filter for smoothing out spikes in the curve but I might need to do something more. Maybe detect spikes algorithmically and dealing them separately. Filtering out the spikes might not compensate perfectly for individual head and headphone resonances but I might get a some sort of average for them. People are going to be affected by the resonance shifts in any case, I just have make sure I'm not amplifying the spikes and dampening the crevices like you said. I also could do different level of filtering for different parts of the frequency spectrum: short window filtering for bass and mids and longer window filtering for highs and then mix the two filters to avoid discontinuities.

Bass cannot boosted too much without introducing distortion. Are there any rule's of thumb how much is safe?

 

[bigshot]

Try it. See if it works.

 

[castleofargh]

Equalizer APO's graphic eq does the impulse responses on the fly, so that's convenient. I'm using measurements from headphone.com and Innerfidelity. Problem with headphone.com measurements is that they have not calibrated their system for Harman target. Fortunately Tyll of Innerfidelity has done so and already published his new compensation curve. I'm thinking of getting all the same models measured by both and finding an average compensation from headphone.com's measurements to Innerfidelity's Harman target compensated curves.

That resonance shift in frequency axis is definitely something to worry about. I remember Tyll mentioning that the spikes in 10 kHz region are caused by ear canal resonances and shift in the frequency axis as headphone position changes. I have Savitzky-Golay filter for smoothing out spikes in the curve but I might need to do something more. Maybe detect spikes algorithmically and dealing them separately. Filtering out the spikes might not compensate perfectly for individual head and headphone resonances but I might get a some sort of average for them. People are going to be affected by the resonance shifts in any case, I just have make sure I'm not amplifying the spikes and dampening the crevices like you said. I also could do different level of filtering for different parts of the frequency spectrum: short window filtering for bass and mids and longer window filtering for highs and then mix the two filters to avoid discontinuities.

Bass cannot boosted too much without introducing distortion. Are there any rule's of thumb how much is safe?

personally I always go easy on the treble Eq for various reasons. even before resonance and potential shifts, there is the issue of trusting a measurement rig above a given frequency. maybe attenuation could be done without much fear, but I certainly go easy on boosts.

for the low end, I guess if increasing the volume level doesn't increase the distortions much, then it's fine to boost. there will come a point where it depends on the listening level but there isn't much you can do about that.
the trouble is how much confidence we can place in low end distortion measurements? Tyll most certainly has his share of graphs where THD+N is more Noise than disto in the low end. his isolation chamber can probably do a good job at mid and high freqs, but stopping a low frequency sound from going through everything isn't easy. I certainly struggle at home with my amateur little rig. it's easy enough to just measure noise and control that it's low enough to consider THD+N being mostly THD. just having both side by side you can control how much the THD+N curve follows the noise one. but just from looking at Innerfidelity graphs, I wouldn't know when to trust those specific graphs.

 

[jaakkopasanen]

personally I always go easy on the treble Eq for various reasons. even before resonance and potential shifts, there is the issue of trusting a measurement rig above a given frequency. maybe attenuation could be done without much fear, but I certainly go easy on boosts.

for the low end, I guess if increasing the volume level doesn't increase the distortions much, then it's fine to boost. there will come a point where it depends on the listening level but there isn't much you can do about that.
the trouble is how much confidence we can place in low end distortion measurements? Tyll most certainly has his share of graphs where THD+N is more Noise than disto in the low end. his isolation chamber can probably do a good job at mid and high freqs, but stopping a low frequency sound from going through everything isn't easy. I certainly struggle at home with my amateur little rig. it's easy enough to just measure noise and control that it's low enough to consider THD+N being mostly THD. just having both side by side you can control how much the THD+N curve follows the noise one. but just from looking at Innerfidelity graphs, I wouldn't know when to trust those specific graphs.

There's two solutions that come to my mind based on your ideas:
1. Scale down equalization on high frequencies. Let's say we have 100% of the damping / amplification up to 6 kHz and then it ramps down to 0% at 8 kHz. So if difference between measurement and target is 6 dB at 6.93 kHz we would have gain of only 3 dB.
2. Ramp down maximum gain for high frequencies. If normally we have selected maximum gain to be 12 dB, it would be at 6.93 kHz only 6 dB and 0 dB at 8 kHz.
Both are quite similar approaches but would have small difference for gains that are below the maximum gain at said frequency. First option scales down gain between 6 and 8 kHz regardless if the requested gain is small or big and the second option would only limit it if the requested gain is above the threshold. 2 dB difference in option 2 would not be affected at all but would be scaled down to 1 dB in option 1 (at 6.93 kHz). Not sure it these make much difference in the sound though.

When it comes to bass distortion it probably also depends a lot on the headphone model in addition to listening volume. Distortion can be measured for each headphone model but, like you said, how much should we trust the measurements? Also Tyll's measurements don't say anything about how much distortion is increased when boosting bass. Listening level is even trickier, there's really nothing we can do about controlling it for the user. So I guess there just has to be multiple options for bass boost levels available and the user can select lower level boost if there seems to be too much distortion. Who knows, some people probably even choose higher bass boost even if it adds distortion just because they like bass so much.

Then a noob question: How are graphs centered around 0 dB axis usually? I've seen at least normalizations about 1 kHz and 200 Hz points as well as some more ambiguous alignments. I was thinking of taking an average difference in 200 Hz to 2 kHz and subtracting that. Is there a norm how this is done?

 

[castleofargh]

yup that's what I poorly implied when I mentioned playback level, there may be an output level that the headphone simply can't handle(physical movement limit) and that indeed is very much dependent on the headphone itself. if you just boost the low end by 10dB while listening to music at 70dB SPL the risk of anything bad going on is obviously small compared to doing the same while listening to the headphone at a level close to what it can handle. in any case, we agree that there isn't much we can do about that. people will use the headphone as loud as they like.

typical alignment is done at 1khz. because it's a cool number, because we're sensitive to that area, or because it's been more or less of a standard. IDK^_^
going for something like 200hz can be interesting for easy calibration of gears because you're unlikely to find massive spikes or significant movements in that area. last I've seen that was for an easy approximation of the SPL output on the miniDSP EARs. it's a good idea because no matter the headphone you use, the 200-500hz area is usually pretty stable all around, the headphone we use, the way we place it, those stuff don't affect that range as much as the rest of the audio band.
I guess it depends on the purpose. to show graphs, 1khz is the usual reference.