But to get this back on track. I will try, with the help of a simple voltage divider and a Creative 0404 USB module, to measure my NAD C325 BEE as soon as possible. I assume I need to generate a tone that locks as closely as possible to a FFT window. A tone of 937.5Hz with a 48kHz samplng rate will be in sync with power of 2 FFT windows.
If I need to dither the signal then I am in trouble, since that will take some studying. But averaging a tone synced to an FFT window is doable quite soon. But for all intents and purposes I wish we could establish a ranking or at least a collection of measurements. Then we could later see which things correlate and which do not. I suggest 1) amp.name, 2) voltage output, 3) load (if any), 4) as many harmonics as possible, as percentage values.
If I need to dither the signal then I am in trouble, since that will take some studying. But averaging a tone synced to an FFT window is doable quite soon. But for all intents and purposes I wish we could establish a ranking or at least a collection of measurements. Then we could later see which things correlate and which do not. I suggest 1) amp.name, 2) voltage output, 3) load (if any), 4) as many harmonics as possible, as percentage values.
There are too many variables when it comes to what is better. So my intent is to just draw attention of proper handling of the interfacing impedance. Will it make a difference in all systems? No. it depends on performance of other parts of the system. But if you have the know how to tackle each link in the audio chain a bit at a time, you will find it rewarding.
Up to now, there is no indication that lower distortion will always result in a better system during listening evaluation. There are different factors that dominate in different systems, solve the most dominant one first, then move on through different parts of the system, and you will get a better total system. If you make a good engineering improvement, but the sound is not right, this usually means other factors are more dominant, and you just have to figure out which one it is. Usually a system can sound pretty bad if you have only a few issues to solve because it's just that combination of problems that make it sound terrible. You will never understand what I say unless you get to that point yourself.
For those who really want to know how much a device alters a signal, have a baseline music file stored digitally, play the music through the device and record the output back to the input. Align the data and generate a difference file. Then you will really know how bad things are.
Last edited:
Thats not what I get. I get that this is 51.2 periods at 48 kHz. It needs to be an exact integer.
960 Hz is exactly 50 periods for a 48 kHz sample. If the time based sampling is locked to the slope of the signal, and then averaged, the 960 Hz tone will average to a steady value while anything that is not in synch with this tone will average down.
The averaging is done in the time domain, not the frequency domain, so the syncronisity must be in time. It is also a good idea to synch the FFT bins as well, so that 1500 Hz would be synched in both time and frequency, as would 750 Hz and 3 kHz. But if you can only synch one, it has to be the time. Otherwise there will not be a SNR improvement or there will be aliasing.
I think the same (or better results can be obtain) by building something like the cordell distorsion magnifier:
CordellAudio.com - The Distortion Magnifier
If you leave out the parts for HF phase matching, the circuit can be build quite simple with a few good quality opamps like the LM4562
Added advantage: you create a buffer between the soundcard and the DUT, which can be usefull when testing DIY-amps...
CordellAudio.com - The Distortion Magnifier
If you leave out the parts for HF phase matching, the circuit can be build quite simple with a few good quality opamps like the LM4562
Added advantage: you create a buffer between the soundcard and the DUT, which can be usefull when testing DIY-amps...
If Earl doesn't mind (which I hope is the case), I could publish a tiny DOS(-Box) executable which does the block averaging on an input file with a block size (any integer) and writing an output file. File format is raw double, array of 64-bit binary double prec. numbers.
Together with Adobe Audition, an excellent (but $$$) audio program, one can do record-while-playback with a 24-bit soundcard (time sync'ed on a per-sample basis) for a block-averaged test.
With Audition, the generation and recording of the test files is also quite simple.
1) Set options to "floating point", and samplerate to your FFT block size, say 65536 (64k), thus giving 1Hz bin width.
2) Generate some minutes of a sine (or a multitone) with only integer frequency(s) but which don't contain 2^N (any N) as a factor, to be choosen to give the expected one(s) when played back and recorded ("multitrack project") at system sample rate, say 48kHz.
3) Dither to 24 bit, play and record, and convert back to float (can be done one the fly)
5) Trim ends and save as .dbl (if saved as .WAV only set/force sample rate but don't resample)
6) Perform Block Averaging
7) Inspect result with a progam than can display one single block of data unwindowed, setting the same block size.
3) Or, use that (or the non block-averaged original recording) to feed DiffMaker, an excellent freeware program for difference analysis.
Of course any repeated stimulus can be used to check for differences and the test can be done with block-averaging before feeding the diff input. When recordings of different amps are made even most of the soundcard's distortion can factored out when a few precautions are made.
Block Averaging may partly filter out or obscure occasional or loosely correlated bursts of noise/distortion etc, something one must consider when judging the results.
Together with Adobe Audition, an excellent (but $$$) audio program, one can do record-while-playback with a 24-bit soundcard (time sync'ed on a per-sample basis) for a block-averaged test.
With Audition, the generation and recording of the test files is also quite simple.
1) Set options to "floating point", and samplerate to your FFT block size, say 65536 (64k), thus giving 1Hz bin width.
2) Generate some minutes of a sine (or a multitone) with only integer frequency(s) but which don't contain 2^N (any N) as a factor, to be choosen to give the expected one(s) when played back and recorded ("multitrack project") at system sample rate, say 48kHz.
3) Dither to 24 bit, play and record, and convert back to float (can be done one the fly)
5) Trim ends and save as .dbl (if saved as .WAV only set/force sample rate but don't resample)
6) Perform Block Averaging
7) Inspect result with a progam than can display one single block of data unwindowed, setting the same block size.
3) Or, use that (or the non block-averaged original recording) to feed DiffMaker, an excellent freeware program for difference analysis.
Of course any repeated stimulus can be used to check for differences and the test can be done with block-averaging before feeding the diff input. When recordings of different amps are made even most of the soundcard's distortion can factored out when a few precautions are made.
Block Averaging may partly filter out or obscure occasional or loosely correlated bursts of noise/distortion etc, something one must consider when judging the results.
free Audacity Audacity: Free Audio Editor and Recorder lets you do much of the same
for short records (by music workstation standards) you can do math on waveforms in MatLab or the free workalikes, SciLab is the one I've used
even Ltspice can be handy for wav generation, inspection, more math options than the audio editor sw
for short records (by music workstation standards) you can do math on waveforms in MatLab or the free workalikes, SciLab is the one I've used
even Ltspice can be handy for wav generation, inspection, more math options than the audio editor sw
The thing is that this technique will not measure below the noise floor of the DUT. My technique will. Thats a huge difference.
I have no problem, but I am afraid that I do not understand what you are doing. Block averaging alone is not the situation. You have to do block averaging on a sinusoid locked and synchronous with the block.
PS Adobe Audition is "Cool Edit", Cool Edit is the better program IMO. You can find Cool Edit around the web. I have a copy from ten years ago or more that I use all the time.
Last edited:
Thanks for explaining, that wasn't clear to me. Do you see an improvement by using both techniques together?
Then take both files, and do a blind ABX comparison (easy possible with the free foobar2000 plugin:
foobar2000: Components Repository - ABX Comparator
Try to get 100% right... Not easy, if you didn't mess up (like clipping /to low levels) during the recording .
Thats possible, but done digitally in the frequency domain, the "nulling" of the fundamental just means setting that bin to zero. Its really rather trivial in DSP. If you did that to the analog signal prior to sampling then you might get some improvment in the sampling errors or numerical calculations, but thats not usually very sigificant.
Basically this is what an old fashion heterodyne analyzer does.
My thinking was that by using the distorsion analyzer the quality of the measuring hardware (ADC or sound card) is less critical, because the distorsion is magnified before conversion. The specs of the LM4562 are much better then most soundscards that can be bought, even very expensive ones. I am however unsure if the above also holds for measuring below the noise floor.. That's why I asked.
Hi Wim,
Yes, it holds. With my mediocre sound card (Waveterminal-192X) and a 1kHz test signal I got the following figures:
Before (digitally) nulling: THD-N = -117.2dB & S/N = 97.3dB
After nulling (i.e. subtracting the distortion from the signal source): THD-N = -140.8dB, so well below the noise level.
I suppose that is what you like to know, right?
Cheers,
E.
Yes, the technqiue will help out a bad sound card. But with proper setup, scaling of the signals etc. even an average sound card is capable of doing the kinds of measurements that I am talking about. Remember, in my test we are looking at the nonlinearities in the DUT as the signal drops, not as it is increased. This means that headroom issues are not relavent and the noise floor issue is strickly that of the DUT and not the sound card. My technique will measure below the noise floor of the DUT which no amount of analog processing will do. Generating a very low distortion and noise signal at very low levels is probably a bigger problem. I do it in code, double precision floating point, so it is very clean.
Yes, thanks. A good reason to continue building the distorsion magnifier...
How do you create a very low distortion and noise signal at very low levels without a pre-amp? My sound cards have only a "digital" volume control, so making a very high resolution signal at very low levels is not possible. When generating a signal at 1Vrms with long averaging (without additional external hardware) I can get good results (see below).
See also the difference between long and short averaging...
Attachments
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.