1. This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
    By continuing to use this site, you are consenting to our use of cookies.

    Dismiss Notice

Headphone CSD waterfall plots

Discussion in 'Headphones (full-size)' started by purrin, Aug 13, 2011.
3 4 5 6 7 8 9 10 11 12
  1. purrin
    OK, here are the "Jazz Mod" results. JaZZ used velvet. Arnaud used foam of varying thicknesses (on the plastic not on the ring). Here, I used one layer of adhesive backed felt on the plastic section at the back of the cup, and two layers on the metal circle surrounding the drivers:
    Again, take these graphs with a grain of salt and know that any comparisons with other measurements (even my own if taken at a later time) would be invalid. I usually take several types of measurements (MLS, sweep, RTA/FFA) and merge the data together. For these graphs, I didn't perform any massaging of the data; but they are good enough for the purposes of this comparison.
    FR graph first:
    HD800 No Mods:
    HD800 w/ mods:
    The red splotches are smaller and the green doesn't push out as much at 5-6KHz. Note the greater uniformity in color horizontally (L-R) across the plot between 0.0 to 0.50ms. This is important.

    Haha! I'm still using an old version with the ISA card on a P3 box. The box is a little on the noisy side and if I expand the time window on the CSDs enough, you can see the whine of the fans! (EDIT: LOL, you can actually see them in the first few waterfall plots in the first post. I've since taken steps to block out this source of noise.) I think the CLIO manufacturer updated it so it's got a USB thingy now.

    I was going to measure these before I sold my rev1s a while back, but I was too lazy. I'm sure someone in the neighborhood has a pair. I would definitely like to measure these.
  2. arnaud Contributor
    Very nice Purrin! As I mentioned in PM, your measurements seem very very clean! For instance, I can't help but think the CSD of the 007+717 shown above has some serious issues (severe aliasing, poor resolution in the time axis too).
    I downloaded a user manual of mlssa and will read up on it later on. But it seems like the basic CSD in that software is simply an fft on the impulse through a moving rectangular window (you'd expect a need for some tapered window though). I am also tempted to try to reproduce the csd analysis using software such as Matlab but - like you - the family comes first and take up quite a bit of the free time :wink:

  3. purrin

    You are giving me too much credit. I'm just happy I was able to find some correlation between the CSDs and what I hear.
    You are right about the CSD. It's just successive FFTs on the impulse response through a window. Don't quote me because I'm not an expert on this, but I believe the end of the window stays the same while the beginning collapses toward the end over successive measurements. This why you see that curve of no data over increasing frequencies over time in post #10. Which of course leads to me to wonder about the validity of the STAX and LCD2 CSDs in post #3. How exactly were those CSD performed? What were the time windows? In what conditions and environment was the CSD measured? Hmmmm.
    "you da man" if you can pull that off in Matlab. I wouldn't even know how to handle the tapering of the window edges.
  4. Anaxilus

    Might be able to scrounge up a local LCD2 rev1.  Let me investigate.
  5. arnaud Contributor

    I don't know, I haven't seen such interesting post in a long while and your results really seem clean. You're probably being too modest :wink:


    Ok, indeed the frequency resolution is directly dependent on the window duration. But here, since we're talking about 5ms or so for maximum window length (to exclude room reflections for speaker measurement in a standard room), this is like 200Hz resolution which is awfully coarse. So I guess one would want to zero pad the end of the window to keep the same resolution over the whole CSD graph. I will read up on this a bit.
    Edit: went through both clio and mlssa user guides and they simply refer to this as the time-frequency uncertainty principle. I believe it just means they're not zero-padding the truncated impulse response so the shorter it gets, the coarser the frequency resolution (like 1000Hz for a 1ms window). Yet again, there is apparently way more than 1000Hz resolution in the graphs so color me confused. 

    The devil is always in the details, so I may bang my head but still want to give it a shot. If possible, would you be able to send me the original impulse response in raw format (csv or such text readable file)? 
    Edit: after going through the user manuals, it really seems to be nothing more than an fft of an impulse response with moving rectangular window (and subsequent smoothing in 1/12 octave bands), which is not far from 10 lines of code in Matlab...

  6. arnaud Contributor
    Purrin was kind enough to send some of his raw data so that I could investigate the issue with his shifting resonances in the CSD graphs. Basically I couldn't reproduce the issue using Matlab software and I suspect it has to do with the post-processing done in that version of Clio software. In particular, as you mentioned Purrin, the time slices don't seem to be zero-padded prior to doing the FFT such that the frequency resolution of the CSD gets coarser with the time (For an fft, the frequency resolution is equal to the inverse of the signal duration in seconds).
    Given the rather short windows, the frequency resolution is actually poor (in the order of 50-200Hz) which is sufficient to make low frequency resonances jump from a 1/12 Octave Band to the next as we go down the waterfall. This is a pretty misleading result (severe frequency aliasing), hence the justification for truncating that data off like you see in stereophile waterfall graphs. The thing is though, if you zero padd the truncated impulse response, you can basically keep the same frequency resolution across the whole time span, I am not sure why these specialized software aren't doing this. Maybe because the recorded data (using MLS) is always too short (few ms) to really see low frequency resonances decay anyway.
    Below are some comparisons of the standard and (Purrin's) modded HD800 headphone...
    Impulse response:  
    Frequency response: 
    Spectral decay - waterfall plot: 
    Spectral decay - contour plot:
  7. purrin


    I not sure whether CLIO version 7 is padding them, but more seriously I think it has a bug. I've run a 2ms window FFT and the results were completely off with the ridge at 10kHz.
    Doh! Great idea, especially for the purposes of this test. On another note, this approach makes think of a better solution for room / speaker measurement: run the impulse response through a function that allows some influence from early reflections, but tails to zero over time, where the 2nd and 3rd reflections from the walls aren't even considered. The bass response would tail off. But you could combine these results with those of another function just for the bass frequencies (where the ear/brain is not going to be able to separate the reflections.) CLIO, at least version 7, only allows a hard gate on the time window. This could save a lot of time because I would't have to resort to sine sweeps or pink noise to measure the bass response.
  8. arnaud Contributor


    I guess the issue with speaker measurements in a room is that the reflections really heavily pollute the response so there's not much solution unfortunately so even just including the first couple of reflections is probably a no no. In the case of headphone though, you could easily treat the walls to prevent reflections at mid-/high-frequencies (see Tyll's test box) and use much longer impulse response to get better CSD graph toward lower frequencies. For the data you sent me yesterday, I noticed the sample rate was 48kHz, I assume there are then limitations on the maximum number of taps on the impulse response result? Anyhow, for headphones, resonances really only start to occur above 1kHz so it's a non-issue I believe. The aliasing of resonance frequencies due to changing freq. resolution across the CSD - if that is really the issue at hand - is a problem though.
  9. purrin
    Here are the FR results for the JaZZ mod with a screen cover strip. Covering part of the screen is supposed to bring out the bass more. I used a one layer strip of dish pack foam.
    FR measurements below. These differences are small, but real. Looks like this mod shelves up by 1 to 1.5db the frequencies below 1.5kHz (I guess we can call this a minor "LCD2 effect" [​IMG].) I went back again to verify that the measurements were repeatable (which they were) because the differences were so slight.
    Green: mod with felt around rings and back plastic
    Black : mod with felt around rings and back plastic + foam strip covering bottom of mesh
  10. purrin

    SET 1 Measurements

    The following were measured in one session. They are comparable to each other. I've got a better more consistent and repeatable method down. Special thanks to Arnaud who pointed out to me some flaws of the CLIO software's output and gave me some better techniques to crunch the numbers.


    The HP1000 is not exactly flat, but this is the better looking CSDs. There is some ringing at 2k, but it's a broad, not sharp ridge. Maybe this is the reason why the Joe Grado does not sound like someone is trying to ram a screwdriver into you ears.





    Dr. Dre Studio Beats


    HD800 no mods


    HD800 felt mods

    Two layers of felt around the ring. One layer of felt on the plastic.

    HD800 Arnaud mods

    Two layers felt around ring. One 2mm foam layer on plastic. Download all three HD800 plots into a picture viewer to see the minute differences among them.
    This is with foam instead of felt on the driver ring.
  11. purrin

    SET 2 Measurements

    This is for the Grado SR60/80 modders:

    SR80i no mods


    SR80i w/ with small piece of Dynamat Extreme on back of driver

    Can't say for certain if the result is within the bounds of measuring error. If the measurements are real, the results are likely not audible.

    SR80i with grill removed (with above mod)

    I was trying to remove just the button, but these seem glued on much better than in past versions, so I just removed the entire plastic grill.
    That "honky" bump at 2KHz is pushed down a bit with the removal of the mesh screen.

    SR80i with inside of outer cup lined with 2mm of foam (with above mods)

    Much smoother at 2kHz than when we started. The foam is made flush with the inner cup.
    CONCLUSION: The SR80i still sounds like an SR80i after these three mods. The sonic signature is too strong for any of these mods to have a significant effect. It would be worth exploring venting the back of the driver for more bass and adding wood cups; but I doubt the severe upper midrange ringing is going to go away.
  12. Anaxilus
    ^  Nice!  I guess it's still not quite an SR225 or RS-1.  Should post this on the mod thread if you haven't.  I remember once saying that according to my ears you won't be able to turn the 60/80 into a 225 or RS-1 w/ a pen and some tack.  I was told to leave the sacred cow alone so I departed Bill's (aka The Lathe) holy temple.
  13. purrin
    Yeah, I got into a really darn good argument with Bill in that thread. I think some head-fiers thought we were going to kill each other until they realized we were actually on good terms and just liked to needle each other. Anyways I would really like to get my hands on some of the better Grado's to measure them.
    I am tending toward my original belief: that while Bill's mods do make a difference, they are not going to make an SR80i sound like an RS1. Ever!
  14. Anaxilus


    I've got some 225 drivers and a set up wood cups staring at each other.  Hmmm.......
  15. purrin


    I did, but I think those guys there now are more interested in how to "bling" an SR80 rather than how to tweak it get better sound.
3 4 5 6 7 8 9 10 11 12

Share This Page