• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Distortion Listening Test

This is an interesting kind of test but, if we put aside the audible noise, we know little about what we can learn out of it absolutely speaking. Only that Mr. X has a better score than Mr. Y.

Klippel knows how to interpret it for their own purpose for sure, knowing what is "real speaker", what the driver choice changes, etc. They shouldn't use it as is though.
 
I could not make the -- test work, the volume was very low, hard to hear anything. For the 6 inch driver option the volume was better. Using DT150 headphones directly on the Mac headphone output.

Skärmavbild 2019-07-23 kl. 00.43.49.png
 
This is an interesting kind of test but, if we put aside the audible noise, we know little about what we can learn out of it absolutely speaking. Only that Mr. X has a better score than Mr. Y.

Klippel knows how to interpret it for their own purpose for sure, knowing what is "real speaker", what the driver choice changes, etc. They shouldn't use it as is though.
There is a paper that goes with it: https://www.klippel.de/fileadmin/kl...rs/Objective_and_Subjective_Evaluation_04.pdf

Alas, it is a very brief paper. It seems that the test simulates the distortions in play for different transducers. They are then using us as a whole to gather data on audibility of it. Sadly seems like scoring is wrong so they are getting false information on threshold of audibility.

The glitch in it is also generating false data.
 
http://www.klippel.de/listeningtest/lt/?page=how

The latter half of this page explains why they are using the 2nd miss as the result reported. It is the median of turning points. Which for someone new to the test looks to be a bad idea from results. It also wouldn't be the 2nd miss if you scored best 1st time, and 3rd best 2nd time and 2nd best 3rd time. They'd go with your 3rd result in that case.

I took it a second time, and managed to get to -39 db before missing, and then -36 and then -39 db. So initial acclimation is making a big difference for your first time taking it. I also didn't hear the noise, but I chose full range speaker instead of "-". There is also a level difference in whether you choose Bl(x) or not.
 
http://www.klippel.de/listeningtest/lt/?page=how

The latter half of this page explains why they are using the 2nd miss as the result reported. It is the median of turning points. Which for someone new to the test looks to be a bad idea from results. It also wouldn't be the 2nd miss if you scored best 1st time, and 3rd best 2nd time and 2nd best 3rd time. They'd go with your 3rd result in that case.
That is a good find. I did guess that they were using the median. But even in that page they don't justify that choice. The test is based on this paper: https://link.springer.com/content/pdf/10.3758/BF03214307.pdf

But that only talks about forward/back-off factors, not taking a median.
 
That is a good find. I did guess that they were using the median. But even in that page they don't justify that choice. The test is based on this paper: https://link.springer.com/content/pdf/10.3758/BF03214307.pdf

But that only talks about forward/back-off factors, not taking a median.
Yes, I was wondering if this came from somewhere else as it doesn't seem to be in that paper. I don't recall this being the norm for up down testing, but I'm no expert in the matter. Paper you link to seems notable for using larger steps with a miss vs smaller steps for a correct choice. Not unusual now, but I suppose it was then.
 
This part from the paper does mention median values:

The psychometric function was assumed to be a tanhlike function. It leveled off at 1.0 for high intensities and at 0.5 for low intensities (see Figure I): p(X) = [e'/(e"+e-") + 1]/2. (2) This could, for example, describe the percentage of correct responses in a 2IFC experiment. The spread of this psychometric function could be defined as X90 - X,so, a difference that would then amount to 1.4. The signal level was set at 2.5 (p = .9967) at the beginning of each track. The first two reversals were discarded. At every even number of reversals, the median intensity of the reversal points was calculated. These so-called midrun estimates were used here as they are used commonly. Any other analysis (e.g., mean of the reversals) would work as well. After 24 further reversals, the track was stopped. Ten thousand tracks were simulated for each condition. The simulation procedure is nearly identical to that used by Kaernbach (in press), which was verified by comparison with experimental data.

And all of this was from a simulation. Not an actual test. Interesting they also discarded the 1st two reversals which would seem to fit with the results people are seeing when doing this test.

I did find a paper that actually tested and confirmed the 5% improvement with 10% fewer trials than equally sized up down testing.

https://link.springer.com/content/pdf/10.3758/BF03334107.pdf

I'd suggest Klippel should disregard the first two reversals, and average the next two.
 
Have not read the background so what does ” real speaker” mean?
 
Have not read the background so what does ” real speaker” mean?
We don't know.
I guess they simulate the distortion occuring at some level on one speaker they are working on.
 
We don't know.
I guess they simulate the distortion occuring at some level on one speaker they are working on.

That would mean a speaker with horrible distortion, which I cannot believe is correct. But if it is the 0 dB reference of the main signal then I can understand that it is in relation to the speaker. For example if the speaker has a distortion of -40 dB at a given level and your detection limit is -33 dB for the sample, you would not hear the distortion from the speaker. But then distortion should be specified at a certain frequency or weighted somehow.
 
@amirm Any chance you can get that clown claiming he can hear -300db down to take this test and crash the statistics with his godly performance?
 
being able to hear distortion products 300 dB below the fundamental?

seriously, the person claiming that needs to get better at lying. he can try taking a look at schiit’s marketing of multibit technology :cool:
 
I got a -39db using LCD-3 phones plugged into an iPad. I understand -40db is equal to 1%, correct? So does that tie to the area that 1% THD is the cutoff for “audible clipping”? (By the way, my tell was fuzziness in the s sounds. I imagine it would be easier to hear in other high frequency sounds like drum kit cymbals. I thought the music recordings were far from reference quality.)
 
Yes, but audible by whom? ;)

By trained listeners, on well recorded songs, I guess. 1% THD seems a bit too high for me, I wouldn't buy a speaker which would have 1%THD, would you?
 
By trained listeners, on well recorded songs, I guess. 1% THD seems a bit too high for me, I wouldn't buy a speaker which would have 1%THD, would you?
Well, I hope I wouldn't, and I hope I would hear the difference. Regarding clipping, I was thinking about amplifiers and their THD at maximum output rather than some average level, which of course should be far lower.
 
By trained listeners, on well recorded songs, I guess. 1% THD seems a bit too high for me, I wouldn't buy a speaker which would have 1%THD, would you?
What speakers do you have, and what is their level of THD at say 95 db?
 
Correct. But audible distortion should be down to -60dB (0.1%) I think.
I believe this is the level people can hear distortion using test tones. Music will vary, but usually distortion will need to be higher for it to be heard.

1% is often used as a guideline for music as a detectable amount. It can vary with the type of music and is around 3% for some types. The number and type of harmonics can effect this cutoff point as well.
 
Back
Top Bottom