[j-curve]

Here I'm going to describe in gruesome, nerd-worthy detail how to make graphs like the ones you can see in the DIY Headphone Response Plots thread. Feel free to tear the methodology to shreds. Many relevant points were raised in a discussion between gerG, halcyon et al. in gerG's thread Want to see a W2002 response plot?

Ingredients
Noise Source. (FM radio with its antenna disconnected).
Audio Equaliser.
Recording Device. (Portable MD recorder, such as the Sony MZ-R909).
Spectrum Analyser. (Spectrogram 5.0 for PC, a fantastic piece of freeware by R.S.Horne).
Playback Device. (Portable MD player, such as the Sharp ST880. Yes, you will need both a player & a recorder).
Binaural In-Ear Microphones (I used Panasonic capsules from a tie-clip mic, details here).
And finally, a bunch of headphones which you can plug into your playback device, in a quiet environment. Background noise will corrupt the data.

Method
First, mix the Noise Sauce.

Connect:-
FM Radio -> Equaliser -> Recording Device -> Spectrum Analyser.
Set the recorder in Rec/Pause mode and equalise the noise to the spectral shape you want. Then record a couple of minutes of Shaped Noise. I thought I wanted white noise (flat spectrum) and that's what I thought I had mixed, but in actual fact I ended up with this:-
Shaped Noise spectrum

Doh! How did that happen? Well, I used the microphone input of my soundcard, which turned out to be MONO with a steep low-cut filter built into it. The filter effect was roughly the inversion of the above graph! Hint: Use the line input of the soundcard, for stereo sound with no stupid filters.

As it happens though, the blunder was fortuitous (Yay!) because the binauaral microphones (and some headphones) have poor low-frequency response, and hence a bit of extra bass in my Shaped Noise was exactly what I needed. In fact, when you use a reference headphone, against which all the others will be measured, it doesn't matter what shape the noise is because that shape will be nullified later. Knowing that, you can tailor the noise to tackle other problems such as low microphone sensitivity, poor signal-to-noise ratio etc. A big smiley noise source with boosted bass & treble would be ideal. Go on... you know your graphic EQ just loves to do that!

Second, record some Headphone Noise.
Slip the binaural microphones into your ears, preferably blocking the ear canals so as not to damage your hearing with the noise. Plug the microphones into the recording device (with a fresh blank disk or tape) and get ready to record some Headphone Noise. Put on some headphones and plug them into the playback device, ready to play the Shaped Noise. Play the Shaped Noise through the headphones at a decent volume and record a sample of Headphone Noise. Repeat for as many headphones as you can find. About 30 seconds of noise per headphone will suffice, except for the reference headphone, for which you will want a couple of minutes. If the headphones have different sensitivities, adjust the volume on the playback device so that the headphones all get tested at a similar loudness.

Third, calibrate the reference headphone.
Connect:-
Playback Device -> Equaliser -> Spectrum Analyser.
Play the recorded Headphone Noise of just the reference headphone and tweak the equaliser until the spectrum measures flat. Use the track-repeat function since this might take a while, and you don't want the player to play back any sounds from the other headphones yet. Once you've got the spectrum flat, don't make any further adjustments to the equaliser(!)

Fourth, graph all of your results!
This is the fun bit.

Simply play back each of the recorded samples of Headphone Noise and watch as the spectrum for each headphone is displayed relative to the reference headphone. Set the Spectrum Analyser to maximum spectrum averaging so as to get a nice, steady and accurate graph.

 

[j-curve]

Ah, before I forget, here's how my reference headphone "flat" spectrum turned out:-
HD580, equalised as reference headphone

(click for HD580 photo)
I used filters as broad as possible, with identical settings in the left & right channels. I didn't want to try and nullify every little bump and dip, especially in the top end, because I figured that would just make all the other headphones look worse without actually improving the accuracy of the experiment.

 

[j-curve]

If you express all the responses in dB then you can add the curves (provided they are independent of each other). The total response should be:-

Shaped_Noise + Headphone + Ear_Flaps + Ear_Canal + Microphone + EQ

Calibrating the reference headphone means adjusting so the total response is zero. If we ignore the ear effects for a moment, this means setting:-

EQ = - (Shaped_Noise + HD580 + Microphone)

Here's what the HD580 plot looked like without EQ (ie. everything in the brackets of the previous equation:-

So the EQ has to be the inverse of this curve. The parametric EQ settings I used were as follows:-
gain @ frequency, bandwidth
+2 dB @ 160 Hz, 2 octaves
-4 dB @ 2 kHz, 2 octaves
-3 dB @ 4 kHz, 3/4 octave
+10dB @ 10kHz, 2/3 octave
+16dB @ 12.5kHz, 1 octave

...continues...

 

[j-curve]

The manufacturer's software indicates that the EQ curve I dialled up is:-

As required, this is the inversion of the raw HD580 plot above, making the total response curve flat for the reference headphone.

Quote:

Number9: No question there is a pattern in their shape for all headphones. A dip about 4Khz and then a sharp rise at 6.5K Hz and then a drop again at 8Khz.

Yes, I noticed the exact same thing, but not for all headphones. Many of them have quite a steep dip starting from about 1.7kHz down to a trough at 3.5kHz. Clearly the EQ curve is not responsible as it is essentially flat there(!) Having the microphones blocking the ear canals could be a factor. On the other hand, there are some notable exceptions to this dip, such as the W100, HD590, CD2000 and K501.

As for the 8kHz diffuse field dip, not all of the graphs show it, but I'm surprised at how many do, including the Z700 and the budget Senns(!)

 

[bangraman]

Edited.

 

[Bobes]

OK, I just got interested in this thread and I have a bunch of questions for you.
1) Are you assuming that since you're normalizing your graphs that the frequency response of the microphone capsules is not a factor in your results? Can you justify this?
2) Do you think that ATRAC compression has an effect on the recorded information, especially below say 15 Hz?
3) What is the effect of the connection mechanism between the source and the sprectrum analyzer? Are you using a digital transfer?
4) If you are using a a digital transfer, does your sound card resample and does this effect your results? Are you considering the inherent noise generated by doing the Analog to digital conversion in the computer?
5) in what sonic environment is the testing done? If your recording binaurally, won't excess input, allowable with open cans, be recorded and impact your results?
6) Are your graphs one-off testings or an average or 3-5 tests per headphone? Can you account for variation between measurements of the same headphone?
Sorry if I seem nit-picky but it's my nature.
Bobes

 

[Joe Bloggs]

You could have gotten white noise much more easily from any of a number of free software waveform generators.

 

[j-curve]

Imperfect Earpad Seal
Some of the plots have major discrepancies in the low end (<1kHz) so I did a little experiment to see what happens when you deliberately break the earpad seal.

HD580 with LEFT earpad lifted slightly
(Cream trace is left)

HD580 with RIGHT earpad lifted slightly
(Blue trace is right)

As you might expect, an imperfect seal can take quite a chunk out of your bass and midrange.

Here's an example from the test results:-

Beyerdynamic DT250-80

In this case it appears that the right earpad wasn't fully sealed, so the white trace gives a better indication of the low-end response of this Beyer.

 

[j-curve]

<a graph>

 

[j-curve]

< a graph>

 

[j-curve]

Bobes, in answer to your questions,
1) Are you assuming that since you're normalizing your graphs that the frequency response of the microphone capsules is not a factor in your results? Can you justify this?
Yes, I do assume that. As far as I know, the microphones rolloff the lows and the highs. The low rolloff is compensated for by the Shaped Noise. The high rolloff is compensated by the EQ Adjustment. The rolloff points of the mics could change according to temperature, pressure, & humidity (ie. the speed of sound) but then so could the response (& sound) of the headphone I'm measuring. If the microphones had any other anomalies in their response then I would have had to compensate for that at the EQ stage, unless the anomaly exactly offset some characteristic of the HD580's response. Either way, the objective of a flat reference spectrum would be achieved.

2) Do you think that ATRAC compression has an effect on the recorded information, especially below say 15 Hz?
Great question about ATRAC since I was surprised that I could record the Shaped Noise and then get the same spectrum back. From what I've read about ATRAC there is a possibility that the algorithm may decide that a certain frequency is inaudible and omit it altogether, so a 20dB dip might become an infinite dip. In our favour is that the energy is distributed widely across the spectrum and psychoacoustics says that widely separated frequencies do not mask each other unless the difference in loudness is substantial. Also, the response curves appear reasonably continuous, with the exception of the 8kHz diffuse field equalisation dip. In short, this could just as easily have been a test of ATRAC, which appears to have passed with flying colours.
Regarding "especially below say 15 Hz?", the minidisc recorder is only rated to 20Hz, and I had to add a disclaimer below 40Hz on the results anyway due to noise problems, so any effects below 15Hz don't concern me.

3) What is the effect of the connection mechanism between the source and the spectrum analyzer? Are you using a digital transfer?
The MD -> EQ -> Soundcard connection is line-level analog. If the connection has any effect on frequency response then this is compensated for by the EQ Adjustment. All Headphone Noise follows the same signal path.

4) If you are using a a digital transfer, does your sound card resample and does this effect your results? Are you considering the inherent noise generated by doing the Analog to digital conversion in the computer?
The A/D conversion seems to be the least of my noise problems. The noise floor of the playback chain at idle is at least 60dB (typical) or 45dB (worst case) below the -30dB line of the graphs, across the audible spectrum. This is well off the chart.

5) In what sonic environment is the testing done? If you're recording binaurally, won't excess input, allowable with open cans, be recorded and impact your results?
Yes, this is the biggest problem I faced. Once I became aware of how serious the background noise could be, I started recording samples of it (with headphones on but no Shaped Noise playing). This allowed me to check that the results were not being contaminated. Some plots had to be redone.

6) Are your graphs one-off testings or an average or 3-5 tests per headphone? Can you account for variation between measurements of the same headphone?
For the most part one-offs, although in the early stages I was testing at least one headphone each time for which I already had data, to confirm that there was some meaning in the results. These had minor differences which could be due to model variation or how I put the mic's in my ears and positioned the headphones on my head. Each recording is actually two tests - the left and right channels. If I get a result in which the two channels agree closely and the background noise was low, that gives me confidence. You make a good point though, since my own HD580 doesn't measure exactly the same as the reference HD580 (although an HD600 was very close). I probably should have invested more time nailing down the reference.

Sorry if I seem nit-picky but it's my nature.
Not at all. Independent critical review can only be good. Thanks for your interest!

Quote:

Joe Bloggs: You could have gotten white noise much more easily from any of a number of free software waveform generators.

True, but I reached for the first thing at hand - the FM radio - which also wasn't going to throw out digital hash like my soundcard. It really was the most remarkably lucky break too, because if I'd used white noise then the first set of results would have been completely worthless from the midrange down. The bass would just have been swamped by noise.

 

[Joe Bloggs]

you could directly burn the software produced wav into CD. As clean as it gets

Also I don't understand what you mean about the bass

 

[j-curve]

Hi Joe,
Unfortunately I don't have CD-R burning technology here... Yep, I'm a Lowtech!

About the bass noise, did you see the bit about the Shaped Noise above? Basically the microphones and headphones roll off at low frequencies and the signal gets swamped by noise, so you need to prepare a special noise source to do this testing. More about noise to follow...

 

[j-curve]

<a graph>

 

[j-curve]

<another graph>