[j-curve]

Noise
Warning: Geek hazard ahead. Please dial a pizza and have your thickest pair of glasses ready.

Here's a plot of the background noise of the MD -> EQ -> Spectrum Analyzer system with the MD in pause mode. Note the different vertical scale of this graph which runs from 0dB down to -90dB, as compared to the test results which are plotted from -30dB to -60dB.

The worst part is below 25Hz and is generated inside the computer. From left to right you can see 50Hz power supply interference coming in on the line input, plus its odd-numbered harmonics at 150, 250, 350 & 450Hz. The little blip at 16.465kHz is probably the horizontal oscillator in the graphics card.

Adding the binaural microphones into the picture raises the noise level. The following recording was done with the microphones sitting on a cushion (ie. not plugged into my ears) in a quiet environment. The recording settings were as follows:-
Mic Sensitivity = Low
Manual Rec Volume = 22/30
These are the same settings as were used for most of the headphone tests.

Now plugging the binaural mics into the ears:-

It seems "body noise" (heartbeat, bloodflow, breathing... hopefully not too much else

) adds about 6dB to the noise floor. Here's how it looks on the normal -30 to -60dB scale:-

Noise Floor

This noise floor represents an absolute minimum for any graph, unless the Shaped Noise can be played louder and the Rec Volume reduced accordingly. The AKG K301 graph is an example of a plot where bass response rolled off sharply and collided with the noise floor. The AKG K240 Monitor was an example where the headphone had low sensitivity and the Rec Volume had to be increased above 22/30. This caused the noise floor to rise about 5dB higher than the above graph.

For these reasons care should be taken when interpreting the low end of the spectrum. One approach is to use the 10-16Hz bucket as an indicator of the noise floor, since the output of most headphones will be down by 20dB or more at those frequencies. Superimposing the noise floor graph over any of the headphone response graphs (and lining them up according to the 10-16Hz bucket) gives a useful indication of whether the noise floor is corrupting the spectrum graph. The noise floor cannot simply be subtracted from the graph though. You might want to put the glasses on now...

Effect of Noise on Spectrum
What is the effect of the noise floor on a response plot? If the noise is in phase with the signal then the two add together, ie. 1 volt of noise plus 1 volt of signal = 2 volts of noise-plus-signal, a gain of 6dB. If the noise is out of phase however, it cancels the signal out completely(!) Unless the noise is correlated with the signal in some way, you can assume that on average it is 90 degrees out of phase with the signal. Under this assumption of orthogonality, a simple rule of thumb can be calculated:-

If the noise floor is 6dB below the signal, the graph rises by 1dB.
If the noise floor is 2dB below the signal, the graph rises by 2dB.
If the noise floor is equal to the signal (0dB), the graph rises by 3dB.

In practice though, we don't know the signal level. All we have is the graph (=signal+noise) and our estimate of where the noise floor is. Applying the rule of thumb in reverse we can develop the "743 Correction" to estimate the true signal level:-

Where the graph is only 7dB above the noise floor, subtract 1dB.
Where the graph is only 4dB above the noise floor, subtract 2dB.
Where the graph is only 3dB above the noise floor, subtract 3dB.

If you apply the 743 correction to each bucket up to about 63Hz you should get a better picture of the low-end response.

Environmental Noise
Noise such as passing traffic, subwoofer demonstrations, air-conditioning systems, escalators, store announcements and a myriad of boombox and home theater equipment on display contributes an ever-changing cacophony to the noise floor. This can be estimated by recording a sample of the environment (while wearing the headphones) and comparing it to the spectral plot with the 743 Correction in mind. In the end, there's no substitute for silence.

Gotta go, my pizza has arrived.

 

[halcyon]

DELETED due to being offensive

 

[j-curve]

Quote:

halcyon: I'm afraid your measurements have several methodological flaws that distort your plots beyond almost anything usable.

I sure hope you’re wrong about that!

I finally managed to get my hands on the journal articles you mentioned. Thanks for the references – they were very interesting! I’d like to make a second attempt at answering the issues you raised, if you will allow… Quote:

halcyon: 1. Test signal is flawed. The proper way is to use short bursts of narrow band (preferably pure) tones. Using constant noise (and even shaped at that) will introduce remarkable errors into the measurements. I know it may be very difficult to use pure (say, 1/3rd octave) tones, but even a sweep would be better imho than a noise. If you use noise, please do use white noise (and do NOT calibrate the headphone to flat FR). Ref: H. Moller, D. Hammershoj, C.B. Jensen, M.F. Sorensen, 'Transfer Characteristics of Headphones measured on human ears', J. Audio Eng. Soc., Vol. 43, No. 4

I thought Moller et al used averaging of repeated impulse responses via an MLS (maximum length sequence) method. That would be nice because it gives phase information in addition to frequency response, but isn’t the test signal equivalent to white noise?

What kind of errors should I be looking for? Intermodulation distortion would obviously be a worry, but there are a few reasons that I don’t think it’s a serious problem here. Firstly, it should be audible (with music that is, not noise

). Secondly, if you believe Sennheiser’s specifications, every one of their current HD range of phones has distortion <0.5%, which is 46dB down. [Is this a valid interpretation? Do those spec’s really mean anything?] Thirdly, many of the graphs do have some big holes in them, which indicates that there isn’t significant IMD, at least at those frequencies. [Small consolation if what you really wanted was sound there!

]
Perhaps there are other effects I’ve overlooked?

My "Shaped Noise" is very close to white noise above 250Hz and pink noise below. As you know, white noise tends to be the technician’s choice, although music averages out more like pink noise. Quote:

halcyon: 2. Recording point well selected, blocking perhaps not Recording point should be at the entrance to the ear canal with a blocked ear canal, but ONLY for headphones that have a free air coupling to surrounding air. Blocked ear canal measurements should not be used with closed headphones as it will distort the data. The blocked ear canals part is something I only dug up myself tonight, so I definitely can't blame you for doing it. I also know that it may be difficult not to block the canals for the duration of the measurement. You'd probably need to have concha level microphones and I can't require you to get those, can I Ref: Ref: H. Moller, D. Hammershoj, C.B. Jensen, M.F. Sorensen, 'Design Criteria for Headphones', J. Audio Eng. Soc., Vol. 43, No. 4

Before I answer that, let’s quickly review Moller’s “FEC” (free-air equivalent coupling) concept, since anyone who hasn’t read the articles probably has no idea what we are talking about. Basically it means that the impedance of air, when looking out into space from the entrance of the ear canal, is almost the same whether you are wearing the headphones or not. The difference is a “PDR” (pressure division ratio) which is plotted against frequency. It is not the same as isolation, although there is probably a strong relationship. Given that just about any headphone changes the way you hear external sounds, I was surprised that Moller’s PDR graphs only showed small variations. Within +/-2dB variation, 5 of the 14 headphones Moller tested were FEC. The winners were something called the BALL (someone post a photo please!) the AKG K1000 (no surprises so far, since neither of those really “connects” to the ears), followed by the Beyer DT990, Stax Lambda Signature and HD560 (this was back in 1995). Moreover, if you allow +/-4dB variations, all of the headphones Moller tested (including the HD250, which is closed), were suitable FEC phones.

Many of the headphones I’ve tested are closed, and no doubt some of them don’t meet the FEC condition, so you make a good point. That could introduce a few dB of give-and-take at some frequency ranges for the non-FEC phones. On the positive side, there are two definite advantages of blocked-ear measurements:-
i) Protects hearing. Shaped Noise is noisy

= pain and possible hearing damage unless you can really do a quick test in a quiet environment.
ii) Less variation between subjects and between measurements. The ear canal adds extra resonances and dips which change from person-to-person. Repeated tests on the same person also have greater variance if the ear canal is open. Hopefully one day someone else will post some real-head results too, and we can do some comparisons. [Hint, hint… It’s easy to do folks, and lots of fun!] Quote:

halcyon: 3. 'Calibration' of reference headphone. The headphone used as a reference and to which all other plots are computed against, is wrongly calibrated to having a flat amplitude response at measurement point. A diffuse field headphone with FEC construction (open) should reproduce something like the below when measure with a blocked ear canal and at the entrance to the ear canal. It should NOT by any means by flat or the sound will be totally unnatural and distorted. [see attached file] Ref: as above (2) Again, to repeat myself. No headphone, whether good or bad, open or closed and whether measured at open or blocked ear canal should measure flat at any measuring position on a listener's head. I think raw graphs without any kind of smoothing or normalisation to a computational 'reference graph' would be the most useful. Why? Because it'll be easier to notice errors from the setup and errors from the headphone itself when looking at raw data. Also, it'll be easier to compare them to the existing 'idealized' headphone measurement curves (such as the one in the attachment, which IMHO is the best out there).

I agree with you about the value of raw data when debugging system problems, but don’t forget that raw response plots are as good as useless for guessing how a headphone will sound unless you have several plotted on the one chart. Even then, you can spend most of your time mentally subtracting one line from another across the whole frequency range.

As a reference point I decided to try and make the HD580 plot flat and have everything else plotted on a dB scale relative to the HD580. Hopefully, people familiar with how a 580 sounds can look at a plot and get some idea of the tonal balance of the headphone in question. Quote:

halcyon: 4. Plotting scale and smoothing improperly chosen The human hearing is at it's most sensitive between c. 2 - 5 kHz. That is why human hearing related measurement graphs plot this range (or some subset of it) with a lot more accuracy than the highest end or the lowest end. I know this is again a limitation of what the software can offer you, but with your current scale some REALLY big problems may be hidden due to the fact that the graphs are smoothed and plotted with such a crude accuracy in the most sentive range of human hearing. Refs: Acoustics and Psychoacoustcs, 2nd. ed., Howard Angus, Focal Press, 2001

If what you want is more space devoted to that area of the graph, then you are right that the software I’m using won’t do that (unless I chop off some of the treble or bass). However, I disagree about the accuracy. The software calculates fast fourier transforms at 4096 points from 10Hz to 22.05kHz with a horizontal resolution of 11Hz. The limiting factor is the graph, which displays exactly 61 pixels between 2kHz and 5kHz, yielding a resolution of 75Hz (at 5kHz). That’s still one pixel for every quarter of a semitone though! Plenty of resolution.
Time averaging is used so that the graph doesn't depend on the exact frequency content of the Shaped Noise over any particular 1/10th of a second, but rather over a couple of seconds. However, there is no frequency smoothing whatsoever.


Sorry for the long post but you made some interesting and controversial points which I wanted to try and understand better. Please let me know your thoughts on the above. I haven’t read much of the literature so I could still be misunderstanding some aspect or other. Once again, thanks for steering me to those articles. Fascinating stuff.

 

[halcyon]

I'm sorry I have offended you.

I tried to provide a reference for each of my argument. I failed.

I have deleted the whole post and I apologize again for constructing a text that was offensive to you.

I stand by my argument and I have provided scientific articles as references for them.

I will not further discuss this matter anymore, unless you specifically ask me to do so (via PM, please). Please do not understand that as anything else except as an attempt to not cause any further hurt feelings.

Please understand that I did not intend to hurt your feelings. I was argumenting against your arguments, not against you as a person.

best regards,
Halcyon

 

[j-curve]

Edit: Removed temporarily.

 

[jona]

j-curve and halcyon-

Thank you both for a WONDERFUL discussion. I learned alot. This thread helps illustrate just how tremendously difficult it is to objectively compare headphones.
Even if some "perfect" lab measurement of the headphone (and/or headphone amp!) is ever achieved, the variability of everyone's unique ear structures will still lead us to choose the cans which are most pleasing to our own ears. And keep us talking about our favorite cans!

Jon

_____

Good discourse is food for the soul.