Electrostatic Headphone Measurements

[arnaud]

I have a strange feeling we're entering a pissing contest on who's got the prettier graph . Nevertheless, I took note of the Torpedo's comment and created yet another version of the post-processing sheet. How about that?

Only worry is that I feel like the HF2 is giving me the finger for showing it naked in one of the graphs. The old (Josepth?) Grado is clearly more behaved

PS: Tyll, eagerly waiting for your data, let's skype on my sat. early morning if possible!

Edited August 31, 2011 by arnaud

[Tyll Hertsens]

Okay, all you math geeks out there, here's a few spreadsheets to play with.

http://www.innerfidelity.com/images/AKGK701.xls

http://www.innerfidelity.com/images/AudezeLCD2Rev2.xls

http://www.innerfidelity.com/images/BeyerdynamicT1SN3964.xls

http://www.innerfidelity.com/images/GradoHF2.xls

Love to see if the results are similar.

[Torpedo]

Now you're talking, arnaud. For my old eyes and brains those look much more readable. Do you have graphs from the LCD-2 or any other orthos? I think it'd be utmost interesting having the waterfall plots of the LCD-2, HD800, 007, 009, Beyer T1, Ultrasone Ed10, and whatever Grado thinks is their top of the line nowadays, GS-1000, PS-1000, RS-1 or all three. The HP1000 look really good IMHO.

[eggil]

If enough interest for testing, I have a pair of Sigma/404 I'd be willing to loan.

[Horio]

If enough interest for testing, I have a pair of Sigma/404 I'd be willing to loan.

You should put them in the spreadsheet (see first post of this thread). Would be great to have some lambda measurements as well.

[Tyll Hertsens]

If enough interest for testing, I have a pair of Sigma/404 I'd be willing to loan.

Sweet, yes, put your name on the spreadsheet. It'll take just a little bit of prep once I get the amp and first pair of cans in, then I'll let you know I'm ready and we'll get them measured.

[purrin]

Thanks Tyll,

Here is quick one of the LCD2r2:

Hard to tell if internal cup - dummy head reflections are playing a part. It's important to note my measurement methods are very different. I'm trying to get an LCD2r1 from someone locally. I'll run the data through with different parameters (-40db scale) at a later time tonight to filter some crap like the 20+ kHz node which appears to be a measurement artifact.

Also, I get the feeling the amplitude in Tyll's data is not pressure in Pa. but db. I've got to re-code to account for this and run the data through again (Yup - just confirmed it - what you now see above is good.)

THIS IS STILL NOT RIGHT - MY FR SCALE IS OFF! ASSUMING A WRONG SAMPLE RATE.

TIME SCALE IS OFF TOO.

I'll let Arnaud run the rest until I back which will be late tonight PST. Arnaud, I think this is good. Took me a while, but this is what I get for hard-coding everything and not documenting my code. Can you confirm with your own analysis?

Edited September 1, 2011 by purrin

[arnaud]

Okay, all you math geeks out there, here's a few spreadsheets to play with.

http://www.innerfide...ges/AKGK701.xls

http://www.innerfide...ezeLCD2Rev2.xls

http://www.innerfide...micT1SN3964.xls

http://www.innerfide...es/GradoHF2.xls

Love to see if the results are similar.

Hi Tyll, is all your data recorded at the same sample rate (170kHz something) and duration (512 samples)?

[arnaud]

THIS IS STILL NOT RIGHT - MY FR SCALE IS OFF! ASSUMING A WRONG SAMPLE RATE.

You've rushed things a bit Purrin it seems . Sample rate is about 171kHz (1/(t(2)-t(1))). One issue I have is that with only 512 samples of raw data, the frequency resolution of the FFT is too coarse to generate a 1/12OB plot (some bands don't have data inside). Might have to do a composite graph (1/3OB until 1kHz, 1/12OB above that). More tonight...

[purrin]

You've rushed things a bit Purrin it seems . Sample rate is about 171kHz (1/(t(2)-t(1))). One issue I have is that with only 512 samples of raw data, the frequency resolution of the FFT is too coarse to generate a 1/12OB plot (some bands don't have data inside). Might have to do a composite graph (1/3OB until 1kHz, 1/12OB above that). More tonight...

Yup. Didn't look at the sample rate, 170357kHz! Exactly on the 512 samples. Too coarse and also not enough to even get any data beyond 3ms which is only good for ~333 Hz upwards.

Edited September 1, 2011 by purrin

[purrin]

Here are a few more on Tyll's data. Arnaud, can you confirm results? I can't help feel that I've messed up on something - I tend to code really quickly without checking the details.

The 512 points would be fine at a lower sampling rate, but at 170kHz, it's just not enough, particularly in the lower midrange. My own HF2 measurements are quite a bit different (they show more ringing). You can see those on HF.

Edited September 1, 2011 by purrin

[Torpedo]

Yep, something is possibly wrong, I've seen other waterfalls of the 701 showing more ringing around 5-6KHz with a decay in the 3-4 ms.

[arnaud]

Sounds like an exponential window was used in the recording to make sure the impulse response fully decayed within the 3ms recording period. I hope this isn't the case as it'll mean the raw data may not be faithfully usable for csd processing...

Edited September 1, 2011 by arnaud

[Tyll Hertsens]

Looks like I may need to change things up a bit. Looking forward to our chat, arnaud, I'll try to read up a bit on what options I have.

[arnaud]

Tyll, the pleasure is mine, you have and are still investing so much of your time for this hobby of ours, I am glad to help in any way I can.

Couldn't get back to the data you posted, probably better to have some understanding of your current data acquisition system anyway and figure out how to best deal with it after. Just in regards to the lack of data in some bands, we can easily zero pad the existing impulse response and make it artificially longer (3x longer should do). That doesn't fix the potential issue with artificial decay (damping) of the exponential window though, so it would be great if you can find out a bit more about any type of "anti-aliasing" windowing being applied during the FRF estimation.

cheers,

arnaud

[purrin]

Sounds like an exponential window was used in the recording to make sure the impulse response fully decayed within the 3ms recording period. I hope this isn't the case as it'll mean the raw data may not be faithfully usable for csd processing...

Arnaud, compare the first 3ms of the impulse response of my HF2 data and Tyll's. It doesn't look like there is a window function to ensure delay, but it does appear there is some high frequency filter function applied after ~1.5ms on Tyll's data. The impulse response starts to look overly smooth with rounded curves. Let me know what you think.

The SR009 is finally shipping in the States, so I should be getting mine sooner rather than later.

Edited September 1, 2011 by purrin

[arnaud]

Edit: error in graphs, deleted

Some good news:

1. I can now automatically load the data off Tyll's spreadsheet (using the columns AK and AL as well as cell B2 in the headphone data sheet) to generate the CSD file (3.5MB per headphone, contains both L & R channels).

2. It will be straightforward to create a batch to process you 100+ files Tyll (will generate a new Excel file with CSD_ prefix or the like)

Here are the results from the data you provided Tyll. I did not spend much time on verifying the validity but the t=0 curve seems to match your FRF graph and the CSD seems along the same line as Purrin's results:

Edited September 4, 2011 by arnaud

[n3rdling]

Those look nice. Is there a reason the graphs start at 500 Hz that I missed?

[Hopstretch]

As an English major, I don't really understand any of this ostentatious sciencing and mathiness. But I am convinced that you gentlemen are doing God's work!

[Voltron]

I am less convinced but really want an explanation or executive summary of this graphy shit.

[arnaud]

Those look nice. Is there a reason the graphs start at 500 Hz that I missed?

One reason is that most bands don't have any results (the resolution of the original data is not sufficient). Another is that there's very little going on below 1kHz for all the headphones I have seen to date. For instance acoustic resonances in the earcup can't really occur below 2kHz because of the size.

For the electrostatic headphones, as mentionned, we may say a lot of resonances from the thin tensioned diaphragm, but I am not sure where it starts. I'd guess pretty low frequency to prevent too much bass roll off.

In any case, it would be good for Tyll to increase the size of his recordings so that we can look at it from lower frequencies and can see it decay fully naturally.

[arnaud]

As an English major, I don't really understand any of this ostentatious sciencing and mathiness. But I am convinced that you gentlemen are doing God's work!

Thanks, you can just enjoy the colors . On my way to add a button to automatically process all the raw data files in the folder, this will really become a "single push button" operation to digest Tyll's 100+ headphones data!

[arnaud]

I am less convinced but really want an explanation or executive summary of this graphy shit.

Simply said, the sharper (frequency) and deeper (decay time) the ridges, the more likely you are to hear these as strong colorations.

[purrin]

Those look nice. Is there a reason the graphs start at 500 Hz that I missed?

Tyll's data has 512 points sampled at 170kHz. At that frequency, the time increment per sample is 1/170k = 0.0000058.

0.0000058 x 512 samples = ~0.003 or 3ms worth of data for the impulse response.

3ms is only just enough time for a signal of 333Hz to complete it's cycle.

Another factor is the nature of transforming the impulse response to a FR - there is much less resolution at lower frequencies than high frequencies. So we may as well start at 500Hz.

I am less convinced but really want an explanation or executive summary of this graphy shit.

One major purpose of CSD's is to find resonances or ringing of the driver (which hurts the ears.) Oftentimes ringing is usually seen on on FR graphs as sharp spikes. But sharp spikes in FR do not always indicate ringing. The CSDs can confirm the ringing. Examples of bad ringing are the John Grado CSDs which have these high ridges throughout the midrange and treble.

Another purpose of these graphs is to see how fast the driver decays. There is some correlation between decay and perceived speed, clarity, and transparency. I just posted a Koss ESP950 measurement at Head-Fi. You can compare this CSD to those of other headphones which are known to sound like shit (Sony XB700, etc.)

These plots are simply tools to confirm observations (and help in the design or modification of headphones, e.g. Fostex modders, etc.), nothing more.

Finally, Arnaud and I are still trying to sort out Tyll's data - to see if it's clean or good enough to crank out meaningful measurements. So it's still too early to give you an executive summary at this time.

Edited September 3, 2011 by purrin

[arnaud]

Edit: error in CSD graphs, deleted

I think I now have a workflow which makes this Excel template usable:

• Can now export all charts automatically to bitmap (4MB file size per chart) or jpg (225kB file size per chart) format
  
• Can now automatically process one or more selected files (in the standard Excel file format you provided us with Tyll)
  

Some example:

Edited September 4, 2011 by arnaud