Really? Can you give us an example of such tiny dB filter ripple in the audio band giving rise to -60dB correlation difference?
But this isn't how the tests on Gearslutsz were created - the analogue output was take directly from the DAC's output - no headphones/speakers or any transducers involved.
What I believe you're saying here is that room responses & reflections will deviate greatly from the input signal - no disagreement here but relevance?
Mark, it gives a ballpark, and no one really cares (nor is there really any motivation *to* care) just *exactly* where that limitation is. Taking Matt to the task for all your frustration about ABX testing is a bit unfair, for what was a pretty off the cuff comment and not a fully fleshed proposal. -40 dB and not worrying about -100 dB gives him a giant fudge factor of a 1000x to work with; let's not lose sight of the forest because of the trees.
As I said, it was for a job... I was working on compression algorithms... Too much close study of the materials looking for artefacts I guess. Sure, probably anyone could train themselves to see these things - you can train yourself to see the frame rate at the cinema* - but why would you do it?! It ruins the enjoyment.
I was helped to learn to ignore the acquired "hypersensitivity" as it were by learning to allow conscious processes to take a sort of back seat. This was some years ago, and was a point of academic interest to a neuroscientist in my extended family who was happy to play with a guinea pig... Probably helps that his missus is a clinical psychologist!
*There's an interesting sideline here, in that some people find high frame rate (48fps) cinema unpleasant to watch. There's some interesting papers on this, I'll try and find....
One could as well as if you could give an example of two data converters measuring at less than -100dB total distortion (maybe less than -120dB), by themselves in one pass giving rise to -60dB nonlinear distortion?
Again, something isn't right here and we don't know what it is. Asking leading questions like a lawyer or politician is more suited to debate than scientific inquiry.
Also, you may know from experience different data conversion anti-alias filters are often audibly different. No surprise there.
Thank you for the reply & the bit I highlighted is exactly what I'm saying is the difference between conscious, analytical listening mostly used in ABX testing & the everyday listening normally engaged in as was stated on page 1 of this thread.
Most people can switch between these two forms of listening & do so in the normal course of auditioning equipment but being required to listen purely analytically as in ABX testing, is difficult without training
Understood. I would invite Matt to rephrase the point he would like to make if he wants. He might have use for making more or less the same point some time again, and maybe this would be a good time to test it out.
That being said, I am not aware of trying to take out frustration in that particular instance. Not impossible though.
First one is easy I took two identical broadband noise signals (~-10dB L and R in the picture). I amplified 1/2 the top one by only .01dB and replaced the top channel (L) with the difference. The R channel is drawn off scale here but the waveform statistics show the difference is identically 0 in the first 1/2 and the .01dB difference in the 2nd 1/2 corresponds to a -59dB null. So +- .01dB simple amplitude variance can account for a ~-60dB null.
My second point is if the difference could be accounted for by benign .01dB or so frequency response differences how would you resolve that from the several dB ripples in any real listening environment?
Attachments
I hope my picture is clear the waveform statistics are on the second half of the signal, the first half is just there for completeness the subtraction yields exactly 0 (perfect null). It's just math BTW, rule of thumb 10% ~=1dB so -60dB = .1% = .01dB.
Last edited:
I am unclear what it is that puzzles you. Unless devices are perfect there will be differences. These differences may be measurable without being audible. Minor wiggles in frequency response are a likely explanation. -60dB difference could mean a 0.009dB response peak; unlikely to be audible?mmerrill99 said:
OK, it seems the weakness of such null testing is that it only shows the worst performance figure irrespective of where that arises from & thus it can be masking audible distortions which are below this headline result . So if the digital output filter is where the largest difference is being recorded - this will define the headline result & every other difference below this is hidden
Probably best illustrated by Archimago example of his use of Diffmaker where both 192kpbs MP3 & NOS outputs give the same corr depth of 50 but sound very different from one another Archimago's Musings: PROTOCOL: [UPDATED] The DiffMaker Audio Composite (DMAC) Test.
Probably best illustrated by Archimago example of his use of Diffmaker where both 192kpbs MP3 & NOS outputs give the same corr depth of 50 but sound very different from one another Archimago's Musings: PROTOCOL: [UPDATED] The DiffMaker Audio Composite (DMAC) Test.
Last edited:
Another vote to try multitone tests, they inherently reject "linear" (still hate this term) distortion which is always at the frequencies in the stimulus and non-linear distortions which are probably more objectionable when present at the same levels.
For this test you might try the standard 32 tone multitone waveform with all clocks synchronously tied to the waveform, then linear and nonlinear distortions are almost completely separable by removing the 32 FFT bins from the output. This is equivalent in some ways to a diffmaker null with no crosstalk between the linear and nonlinear effects.
Last edited:
Sure, can you give us the typical range of IMD levels you found with your multitone tests on DACs without naming any - just a typical range ?
Edit: I see you edited after I posted. Still interested in the typical range of IMD you found
When you say all clock synchronously tied to the waveform, can you explain what "tied to the waveform" means?
Last edited:
Sorry I didn't compile the results in terms of total rms noise floor so right now I would be speculating wildly. I'm willing to give a Scarlett 2i2 a try for at least one data point.
By synchronous I mean the exact same clock for the DAC and A/D, when you use for instance two different USB sound devices with separate clocks the tones are no longer exactly in 1 FFT bin (usually) and in gross cases they can be way off. For instance I have a sound card that recorded a 1kHz sine from a lab standard generator as 1006 Hz but only at 44.1 sampling rate. The clock divider was flawed. Just another point that as Mark says you have to question results when on the face of it they make no sense. Remember the data sheets for DAC's and A/D's stress the non-linear distortions only so when they both reproduce an output to input at -100dB levels the -60dB null makes no sense. It's entirely possible that the GS tests are of no practical use, this must be on the table.
EDIT - By tied to the waveform I mean you pick frequencies that are an exact multiple of the sampling frequency. I simply prefer to use 96K FFT's so every bin is exactly in Hz.
By synchronous I mean the exact same clock for the DAC and A/D, when you use for instance two different USB sound devices with separate clocks the tones are no longer exactly in 1 FFT bin (usually) and in gross cases they can be way off. For instance I have a sound card that recorded a 1kHz sine from a lab standard generator as 1006 Hz but only at 44.1 sampling rate. The clock divider was flawed. Just another point that as Mark says you have to question results when on the face of it they make no sense. Remember the data sheets for DAC's and A/D's stress the non-linear distortions only so when they both reproduce an output to input at -100dB levels the -60dB null makes no sense. It's entirely possible that the GS tests are of no practical use, this must be on the table.
EDIT - By tied to the waveform I mean you pick frequencies that are an exact multiple of the sampling frequency. I simply prefer to use 96K FFT's so every bin is exactly in Hz.
Last edited:
Thanks
Yea, I got the synchronous bit, I just didn't know if "tied to waveform" was significant?
Yea, I got the synchronous bit, I just didn't know if "tied to waveform" was significant?
Ok, I see where you are going, but no, what I'm saying does not support your argument. Rather different cases.
OK but this gets complicated - don't the tone frequencies need to be arranged so that their IMD products don't fall on other tones & therefore are hidden? Sorry if this is OT
It time aligns as best it can - is this based on some sample period of analysis between the two tracks? If the tracks drift in time during the sample being analyzed does it dynamically align them? This would be wrong IMO as it could also hide other timing differences that are not due to clock drift.
I reckon it's probably best to use synchronous clocking as Scott says & turn off timing compensation in Diffmaker
Last edited:
The blurb on the Diffmaker download page leads me to believe it compensates for frequency response anomalies.
Under the heading 'What can Audio Diffmaker do?' it says, inter alia :
Measure the frequency response of the equipment being tested and apply it so the effects of linear frequency response can be removed from the testing.