[upstateguy]

Quote:

Originally Posted by nick_charles /img/forum/go_quote.gif Okay, a digital recording using this system is more accurate in terms of amplitude for frequencies where the input signal level is higher. So for the dominant frequencies which are about 3Khz in this example ( a set of cymbal crashes) the input signal levels are about -28db and the recording shows miniscule deviations from this (0.000071db). As the input signal level decreases the recording becomes (relatively) less accurate. So for frequencies with less energy the recording is less accurate. The practical upshot is that until the signal level drops below -73db the recording is pretty accurate in terms of amplitude (never deviating by more than 0.01db and rarely that much) , but as the level drops more variation creeps in so that for frequencies with a signal level of less than -108db the recording deviates by up to 0.4db. But, there are only 12 frequency points at which the difference between input and recording exceeds 0.1db and this is out of 32768 samples and only 43 where it exceeds 0.05db and all of these happen for frequencies with an amplitude of less than -88db . For only 2% of all frequencies tested does the deviation reach 0.01db.

Is any of this stuff audible?

USG

 

[leeperry]

Quote:

Originally Posted by nick_charles /img/forum/go_quote.gif Okay, I tried Wavespectra and Sinegen. Wave spectra is fun but it does not seem to export the results, I am also not quite sure I can interpret the results as it showed a THD+N of 700% ?? Does not seem to help for what I want to do, but thanks anyway fun to play with

ahh, maybe 20.5kHz is a tad much...I only used it w/ 1 and 10kHz in 16/48 stereo.

ideally you should get the exact same figures for all your samples, and I'm quite sure you won't.

 

[nick_charles]

Quote:

Originally Posted by upstateguy /img/forum/go_quote.gif Is any of this stuff audible? USG

According to the body of prior psychophysics research these differences in the context of the specific files analyzed should not be audible. For instance two stimuli at 20K differing by 0.49 db might be marginally differentiable at absurdly high levels for someone with superlative hearing and in isolation but at -108db and embedded in music not a chance.

That is the theory anyway, but the samples are out there and anyone can try a DBT on them and prove theory lacking...

 

[nick_charles]

Quote:

Originally Posted by leeperry /img/forum/go_quote.gif ahh, maybe 20.5kHz is a tad much...I only used it w/ 1 and 10kHz in 16/48 stereo. ideally you should get the exact same figures for all your samples, and I'm quite sure you won't.

Well, we already know that the reference and the recorded samples are not identical and that there is variation between the different samples, we established this on page 5.

Nevertheless I will run a music sample and the reference through it and have a look at the spot noise on the dominant frequency for the two...

Hmmm, I ran the reference and a recorded sample through and did a spot test at 3465.2hz , the dominant frequency at sample number 5000, the reference showed a max of -30.97db and the recorded sample a max of -31.11db, actually all the recorded samples showed the same max at the peak frequency, which is a bit suspicious, then I loaded up the reference file again and it now showed a max of -31.11db, at least on my system under Vista it seems a bit capricious ? THD shows consistently in the region of 54%

 

[leeperry]

well, that'd give a rough idea of the actual harmonic distortion going on...say for a 10kHz sine test tone.

 

[leeperry]

ok, I've put a sample here for your consideration: 20.5kHz.rar



a wild guess would be that THD will not be 0% anymore? I can also provide 1/10/15kHz test tones.

 

[upstateguy]

Quote:

Originally Posted by leeperry /img/forum/go_quote.gif ok, I've put a sample here for your consideration: 20.5kHz.rar a wild guess would be that THD will not be 0% anymore? I can also provide 1/10/15kHz test tones.

I could only hear it if the volume was extremely high....

What is the significance of this WAV?

USG

 

[leeperry]

to record it over toslink, and see if the THD/THD+N/SNR figures are any different...and how the FFT size/window functions interact w/ the differences.

 

[nick_charles]

Quote:

Originally Posted by leeperry /img/forum/go_quote.gif well, that'd give a rough idea of the actual harmonic distortion going on...say for a 10kHz sine test tone.

It is an interesting product and I ran some tests on my files, the results in ters of levels were consistemt with Cool Edit Pro, but you have to click on each peak individually to get any data, so it is best suited to spot tests rather than getting a good overall picture.

 

[upstateguy]

Quote:

Originally Posted by leeperry /img/forum/go_quote.gif to record it over toslink, and see if the THD/THD+N/SNR figures are any different...and how the FFT size/window functions interact w/ the differences.

IC, thanks.

Edit to ask how many db it's down?

 

[nick_charles]

From a bit of experimentation, following what Bohdi said it seems you get the best analytical results from a FFT size that is exactly the same as the number of samples. For instance with a 64K FFT and a 1.4860779 second sample and a hanning window you get the narrowest skirt on a 1K 44,100 sample. If you use 128K samples on the same file you get misleading results, this might explain the 540% THD

.


Similarly a 44,100hz, 0.743038 second sample and a 32K FFT gives better results than a 64K fft !

So I need to alter my samples and burn a new test CD

 

[leeperry]

Quote:

Originally Posted by nick_charles /img/forum/go_quote.gif it is best suited to spot tests

AFAIK, that's how all the THD/THD+N measurements are made...w/ a 1kHz sine test tone?

maybe 20kHz would be wiser, as this is where you got the highest difference so far.

 

[b0dhi]

Just in case anyone's using Wavosaur for any of this analysis - when I was testing RMAA, I noticed that THD increase by 300% after I did something with Wavosaur. So I did some more testing and found that simply opening the RMAA file and re-saving it increases RMAA's THD measurement by 300%. I tested with Adobe Audition and it didn't have this problem. Using Audacity, the THD increased by 100%. I assume/hope it's dithering causing this.

Also, Nick_charles, could you please upload some of the reference files you used? I.e., a couple of files that were passed through your digital chain? I'm getting some large diff results on those ABX files and want to make sure it's not noise or something.

Filtered to only show differences for spectra louder than -70dB:



FFT of the Diff above:



I also did the transform using a Flattop window and got similar results.

 

[nick_charles]

Quote:

Originally Posted by b0dhi /img/forum/go_quote.gif Just in case anyone's using Wavosaur for any of this analysis - when I was testing RMAA, I noticed that THD increase by 300% after I did something with Wavosaur. So I did some more testing and found that simply opening the RMAA file and re-saving it increases RMAA's THD measurement by 300%. I tested with Adobe Audition and it didn't have this problem. Using Audacity, the THD increased by 100%. I assume/hope it's dithering causing this. Also, Nick_charles, could you please upload some of the reference files you used? I.e., a couple of files that were passed through your digital chain? I'm getting some large diff results on those ABX files and want to make sure it's not noise or something. Filtered to only show differences for spectra louder than -70dB: FFT of the Diff above: I also did the transform using a Flattop window and got similar results.

Here is the reference file a 3.6 second cymbal crash

07-130 chinese cymbal loop.wav - DivShare

Below are two recordings from CD


T6

T5

But as I said in a earlier post, and as you suggested, I think the reference file is too big, I am going to crop them to fit either a 32K or 64K FFT exactly as this seems to give more believable results. Nevertheless, when I did an invert in CEP and added the inverted file to a recording so that the only thing left was the difference I got a difference file where the highest value was -120db - so something is a bit screwy here , the end reslut was a straight line with some small deviations as expected showing in the graph I posted.

 

[b0dhi]

Thanks Nick_charles.

To be honest I think what you suggested there, i.e., taking the difference of the files, is the best approach. I'm not at all convinced an FFT is a reliable way of detecting differences in these files since the margin of error seems to be too high. I'll see if these files give more believable results though.

Edit: Looks like the differences are there in those reference files too. These are a diff for t5 and t6. Assuming those two files are just recordings of the same audio through the same audio path, I think it's safe to say these differences are due to FFT "error". These are for maxima louder than -70dB by the way.


And FFT for the above diff:


Edit2: did a diff between two of the same file and got precisely zero difference. Then I did a diff between the Blackman-Harris and Flattop transform of the same file and got differences in the order of 50db! Something isn't right here. Similar results with Blackman-Harris vs Kaiser-Bessel.