[Joe Bloggs]

Weird...wonder if that's accurate....over 10ms+ decay in the lower midrange and below? Maybe I don't know how to read graphs.

An inherent flaw of these cumulative spectral decay graphs is that they are based on the raw measured frequency response of the earphones and are greatly influenced by this frequency response. For example, big long decay in the bass for any earphones with elevated bass response according to raw measurements. Treble peaks that almost always turn into ridges running down the graph. (Though not really in this case, for some reason). At my suggestion, xnor once applied minimum phase EQ to cancel out the effects of frequency response on a CSD plot of a pair of (unknown, but reportedly not really HiFi) earphones. That discussion ran from p. 49 to p.51 in this thread: http://www.head-fi.org/t/566929/headphone-csd-waterfall-plots/720#post_8475205

The discussion can be summed up in two graphs. One the raw CSD:

One the same CSD compensated with a minimum phase EQ filter to cancel out the effects of frequency response:

When compensated, the result was hardly recognizable as a CSD plot, almost all the "ringing" and "decay" was gone, this showed that (pulling a number out of my ass) 90% of all the ringing and decay you see in usual CSD plots are simply the direct result of the earphones' frequency response and not really indicative of any time domain aberrations inherent in the driver / housing...

And until compensated CSD plots become the norm, I wouldn't look into CSD plots for any insight into the sound of any headphones...

 

[ohotonge]

An inherent flaw of these cumulative spectral decay graphs is that they are based on the raw measured frequency response of the earphones and are greatly influenced by this frequency response. For example, big long decay in the bass for any earphones with elevated bass response according to raw measurements. Treble peaks that almost always turn into ridges running down the graph. (Though not really in this case, for some reason). At my suggestion, @xnor once applied minimum phase EQ to cancel out the effects of frequency response on a CSD plot of a pair of (unknown, but reportedly not really HiFi) earphones. That discussion ran from p. 49 to p.51 in this thread: http://www.head-fi.org/t/566929/headphone-csd-waterfall-plots/720#post_8475205

The discussion can be summed up in two graphs. One the raw CSD:

One the same CSD compensated with a minimum phase EQ filter to cancel out the effects of frequency response:

When compensated, the result was hardly recognizable as a CSD plot, almost all the "ringing" and "decay" was gone, this showed that (pulling a number out of my ass) 90% of all the ringing and decay you see in usual CSD plots are simply the direct result of the earphones' frequency response and not really indicative of any time domain aberrations inherent in the driver / housing...

And until compensated CSD plots become the norm, I wouldn't look into CSD plots for any insight into the sound of any headphones...

well said, actually I reluctantly included the csd graph because usually it doesn't provide any extra info that fr graph already provided and can cause wrong impression.

 

[ohotonge]

You could call it extension as well. The sound made after the strike (the shhhhhhh until the end).

thanks for clarification. but still, I don't think the extent of 'decay' can provide fr extension info.

 

[arnaud]

I am a bit confused by the interpration of what xnor did:
- i followed that thread I recall because I was looking at something similar at the time (that is remove measuring head signature from the temporal measurements including impulse response and related csd)
- now, what xnor did is prove that headphones pretty mucj are minimum phase systems so that 1) you can reconstruct the phase from an FRF magnitude response and 2) (what xnor did) create an inverse minimum phase filter that effectively gives the phone a flat amplitude response and short decay by the same reason...
- Obviously, if you damp a resonance to such amount it no longer shows up in the frequency response (nor its impulse response because the inverse filter is correct in both amplitude AND phase), it won't show up the csd plot either

Now, in regards to long ringing at low frequency:
- For the same Q factor (how peaky a resonance is), a lower resonance frequency has comparatively much longer decay time than a higher frequency resonance
- think of it as a mass on top of spring and dashpot. The spring and mass govern frequency of resonance, the dashpot how peaky the resonance is (Q factor)
- in frequency response chart this looks like a resonance with more or less wide bandwidth
- in time domain, if you just strike the mass once and let it oscillate until it goes to rest it will take a few cycles depending on how effective the dashpot it (underdamped means more than one period to stop, overdamped less than one period)
- one oscillation period is the inverse of the resonance frequency. Hence, the higher the resonance frequency, the shorter the oscillation period (in seconds)
- that is it: for same damping level (same number of oscillations to get bacl to rest), you get much longer "ringing" for a low frequency resonance...
- now maybe what joebloggs was suggesting is that because not all frequencies have the same amplitude at the 0ms step, naturally they won't take the same time to drop in level by say 20dB. This is true but you can then simply normalize the csd graph by the 0ms amplitude and there absolutely no way you'll get the same result as what xnor did

Now what does it tell me about the ks1500 measurement:
- This might be an artifact of the csd processing (in particular the ringing of the window used for truncating the impulse response)
- however, look at both Frequency response and distortion figure, seems to me like the earphone is tuned to have a resonance around 100Hz region
- There's another nasty one at 10kHz but that could just be acoustic resonance of the coupler so not sure it's worth interpreting
- About comparison in high frequency extension with se846, I'd tend to agree with ohotonge that some could misinterpret a less damped resonance at 10kHz with additional treble extension when hearing cymbal shimmer more clearly

Cheers,
Arnaud

 

[arnaud]

An animation is worth a thousand words :
- see figure 8 for the animation in time domain
- see figures 15 and 16 for the corresponding frequency response function

http://www.maplesoft.com/applications/view.aspx?SID=3926&view=html

 

[goodvibes]

An inherent flaw of these cumulative spectral decay graphs is that they are based on the raw measured frequency response of the earphones and are greatly influenced by this frequency response. For example, big long decay in the bass for any earphones with elevated bass response according to raw measurements. Treble peaks that almost always turn into ridges running down the graph. (Though not really in this case, for some reason). At my suggestion, @xnor once applied minimum phase EQ to cancel out the effects of frequency response on a CSD plot of a pair of (unknown, but reportedly not really HiFi) earphones. That discussion ran from p. 49 to p.51 in this thread: http://www.head-fi.org/t/566929/headphone-csd-waterfall-plots/720#post_8475205

The discussion can be summed up in two graphs. One the raw CSD:

One the same CSD compensated with a minimum phase EQ filter to cancel out the effects of frequency response:

When compensated, the result was hardly recognizable as a CSD plot, almost all the "ringing" and "decay" was gone, this showed that (pulling a number out of my ass) 90% of all the ringing and decay you see in usual CSD plots are simply the direct result of the earphones' frequency response and not really indicative of any time domain aberrations inherent in the driver / housing...

And until compensated CSD plots become the norm, I wouldn't look into CSD plots for any insight into the sound of any headphones...

And the same holds true for distortion curves. You'll often see bass distortion at well over 1% on IEMs that have elevated bass response because it's referenced to 1khz amplitude. Drop that to linear and the distortion % falls accordingly. 
 
As for the frequency response. That wouldn't bother me at all if it sounded right to me. Some roll is needed to sound natural and the rest is iffy in IEM measurements. The device is certainly capable of extended response.

 

[Joe Bloggs]

I am a bit confused by the interpration of what xnor did:
- i followed that thread I recall because I was looking at something similar at the time (that is remove measuring head signature from the temporal measurements including impulse response and related csd)
- now, what xnor did is prove that headphones pretty mucj are minimum phase systems so that 1) you can reconstruct the phase from an FRF magnitude response and 2) (what xnor did) create an inverse minimum phase filter that effectively gives the phone a flat amplitude response and short decay by the same reason...
- Obviously, if you damp a resonance to such amount it no longer shows up in the frequency response (nor its impulse response because the inverse filter is correct in both amplitude AND phase), it won't show up the csd plot either

Not obvious, because not all systems are minimum phase. If a pair of earphones is totally minimum phase, it seems silly to me to talk about how much ringing it has at such and such a frequency for how long when simply tweaking an equalizer to reshape the phones' frequency response will also totally eliminate the ringing / decay problems. But I suppose this may not be relevant to those who don't equalize.

Now, in regards to long ringing at low frequency:
- For the same Q factor (how peaky a resonance is), a lower resonance frequency has comparatively much longer decay time than a higher frequency resonance
- think of it as a mass on top of spring and dashpot. The spring and mass govern frequency of resonance, the dashpot how peaky the resonance is (Q factor)
- in frequency response chart this looks like a resonance with more or less wide bandwidth
- in time domain, if you just strike the mass once and let it oscillate until it goes to rest it will take a few cycles depending on how effective the dashpot it (underdamped means more than one period to stop, overdamped less than one period)
- one oscillation period is the inverse of the resonance frequency. Hence, the higher the resonance frequency, the shorter the oscillation period (in seconds)
- that is it: for same damping level (same number of oscillations to get bacl to rest), you get much longer "ringing" for a low frequency resonance...

Again, this analysis only holds for a minimum phase system. A mixed phase system would not behave in such a tidy manner.

- now maybe what joebloggs was suggesting is that because not all frequencies have the same amplitude at the 0ms step, naturally they won't take the same time to drop in level by say 20dB. This is true but you can then simply normalize the csd graph by the 0ms amplitude and there absolutely no way you'll get the same result as what xnor did

That's just it though. Intuitively you'd expect that equalizing the impulse used for the CSD graph would just affect the starting height and not the rate of decay. What xnor showed was that if you adjusted using a minimum phase equalizer (which is very common), the decay of all frequencies actually sped up drastically.

All this ought to be more relevant to KSE1500 owners than others, because there's a 4-band parametric EQ built right into the system.

 

[arnaud]

JoeBloggs:

As I understand, headphones are actually minimum phase systems (mechanical, acoustic, electrical resonances and the mixture it contains). Do you have data showing otherwise as I am very curious about this one.

About the equalizing using minimum phase filters, it is apparently easily doable and xnor proved. Main issue is what are you shooting for because what you're looking at with these raw data measurements are just as much the headphone's response as it is the coupler influence, even for something as deceptively simple as an iem coupler...

Why digital eq is not used much to date I think is just as much remnants of godawful 70´s era analog filters (e.g. EQ and audiophile typically don't go in pair) as the impracticality of it (few companies are selling headphones with in line DSP and converter / amp ...).

As to your last point, I still believe you're confused (or I don't understand you) in regards to two very different things:
1. Computing a csd on an raw impulse response and normalizing the graph by the amplitude at t=0 (something that everyone should do, and that indeed also applies to distortion figures)
2. Applying some correction filters to a headphone (inverse based of magnitude response using minimum phase filters as xnor did for example) which effectively converts a) the phone's impulse response into a dirac , b) its frequency response function into a straight line, c) its csd into a very sharp wall (that's the idea with a dirac function....)

Cheers,
Arnaud

 

[arnaud]

Getting a bit tangent to the topic at hand but:
- non-minimum phase system means it's inverse is either non causal and / or unstable
- practically speaking, that means if we were to measure a very deep trough in the frequency response and try to eq it (setting a very sharp resonance right at that frequency)
- such very narrow drop in response may be a measurement artifact (microphone sitting in a nodal line) and also much less intrusive than a resonance so typically would not need equalizing
- as such, I guess even if there may be non-minimum phase parts of a given measured headphone response, it is something the educated eye would ignore (or otherwise smooth the response curve so it's isn't quite visible).
- I could not find anything about headphones but people have already looked into minimum-phase characterics (or lack thereof) of speakers in regular room: http://www.roomeqwizard.com/help/help_en-GB/html/minimumphase.html

Cheers,
Arnaud

 

[goodvibes]

It's not entirely that simple. Dropping amplitude isn't enough to avoid ringing. EQing it down lessens the effect but true resonance can't be defeated with EQ. The mean amplitude will be correct and show tidy but dynamically the initial pulse or arrival will be low compared to other frequencies. It's not a cure all and more a bandiad if at times an effective one. if it's simply frequency and less so resonance, EQ is a cure but more often than not resonance is in the mix. Same reason room correction often sounds better but is never as good as a good room. Resonance and response are somewhat interactive but not the same thing. Ears/brain are probably more sensitive to time relationships than anything else. That said, I agree with Joe about how FR affects those plots and about housings in general. Housings of IEMs are way overrated as they are more about holding the driver in place and orifices than anything else. The surfaces are just too small and light to have any sort of significant resonant behavior. I would think there's more effect from tip flex than any resonant behavior from an rigid IEM surface. 

 

[Mimouille]

It's not entirely that simple. Dropping amplitude isn't enough to avoid ringing. EQing it down lessens the effect but true resonance can't be defeated with EQ. The mean amplitude will be correct and show tidy but dynamically the initial pulse or arrival will be low compared to other frequencies. It's not a cure all and more a bandiad if at times an effective one. if it's simply frequency and less so resonance, EQ is a cure but more often than not resonance is in the mix. Same reason room correction often sounds better but is never as good as a good room. Resonance and response are somewhat interactive but not the same thing. Ears/brain are probably more sensitive to time relationships than anything else.

Are you sure? Because I always thought that ringing could be avoided by...just kidding, no clue what you guys are talking about

 

[spook76]

Are you sure? Because I always thought that ringing could be avoided by...just kidding, no clue what you guys are talking about

I am with Michael, what is the sum of your discussion? Are graphs provided accurate and reliable?

 

[csglinux]

 
 

 

https://www.seeko.co.kr/zboard4/zboard.php?id=m_phones&page=1&sn1=&divpage=1&sn=off&ss=on&sc=off&select_arrange=headnum&desc=asc&no=144

Interesting post ohotonge - thanks for sharing this! This prompted me to pull out my tone generator last night and have a listen. There's definitely a bump between 4 and 10 kHz. Hard to tell by ear, but it could easily be around 10 dB. BTW, your website is a great resource. I just wish my Korean was a bit better I noticed that those IEMs I'm familiar with (UE900s, SE846) seem to have a higher peak ~ 10 kHz and a sharper roll-off than I've seen measured elsewhere (e.g., Innerfidelity). I noticed you use 24th octave band bins, whereas Innerfidelity don't specify - I'm guessing they use PSD? I'm curious as to what might cause some of this gray area beyond (may even around?) 10 kHz. Do you use any kind of windowing functions or sliding averages in your FFTs? How do you do your diffuse-field corrections? How good is your mic and do you make any corrections for it? What type of eartips were used on the dummy head and how do you ensure a good seal?
 
Joe, I think I understand your comments, but isn't a resonance - and any resulting extended decay - from anything (be it driver or housing) potentially an aberration? You can't use the driver out of the housing, so whatever the cause, those sound waves are going to hit your eardrum the same way. I could potentially tame resonances a bit using that nice parametric EQ, but since discovering SpinFits, I haven't bothered to. I actually like the FR of the KSE1500 as is. Depending on measurement, FFT settings & interpretations we might debate that 10 kHz peak somewhat, but in any event I think ohotonge's measurements show that calling the KSE1500 "neutral" or having a "rich neutrality" might not be entirely accurate  I have no problem with that. I've always felt a V-shape is needed, especially for sound-isolating IEMs where the intention is to drive lower SPLs where our ears aren't as sensitive to the frequency extremes. I'll probably be banned for saying this on head-fi, but I still love my ATH-M50x with their rather peaky treble. I guess YMMV at higher volumes, but I never go there.
 
I'd be interested to see if these measurements can be repeated elsewhere.

 

[goodvibes]

The graphic is good but we can hear a lot more than 36 db down from mean. If it's resonance and the graph is expanded down, you'll still see it though greatly reduced. It's been manually cut by about 15 db but the overall result is much better than that. There are no free lunches as I haven't found an EQ that isn't at least a little audible and things behave a bit differently in dynamic situations than in sustained sweeps.

 

[csglinux]

  The graphic is good

Not to disrespect or impune ohotonge, but there are always uncertainties in measurements. Even with accurate measurements, there are a zillion ways to compute and present FFTs from time-domain data. Check out those headphone measurements common to both seeko.co.kr and innerfidelity. There are some significant differences, particularly in the higher-frequency resonance peaks that cannot, IMHO, be purely attributed to the use of 24th octave-band bin averaging. They can't both be right. These graphs might be good, but there's no way of knowing that just by looking at them.