 
|
Audio Asylum Thread Printer Get a view of an entire thread on one page |
For Sale Ads |
|
|
Audio Asylum Thread Printer Get a view of an entire thread on one page |
For Sale Ads |
193.150.10.5
In Reply to: RE: TESTS: Round-up of Windows and Mac Audio Players. posted by Ugly on June 15, 2013 at 11:34:52
Hi,
> Interesting. I'm not familiar with the software but had assumed
> it would be somehow correlated to reality.
It correlates (within the limits of the device used as A2D) quite well with reality, however you must understand how FFT works and how FFT noise floor relates to a single figure SNR.
In other words, you must understand the subject sufficiently and how to interpret the output of the test system. Please read up on "FFT process gain" to understand why an appx. -133dBFS FFT noise floor is only 16 Bit equivalent.
Ciao T
Sometimes I'd like to be the water
sometimes shallow, sometimes wild.
Born high in the mountains,
even the seas would be mine.
(Translated from the song "Aus der ferne" by City)
Follow Ups:
Based on your post I searched and found the linked paper. It addresses process gain at the end of the article.
If I've understood the concept it suggests the 24bit ADC in the E-Mu 0404 USB Archimago claims to use should have a theoretical quantization noise floor limit near -146dB. With a process gain of nearly 70dB the FFT noise floor is somewhere around -216dB per the formulas given.
I'm sorry if I'm missing the obvious but how is this concept of process gain accounting for what you are saying.
Hi,
You are way off track. As you are already looking at an FFT you need to SUBTRACT (not add) process gain.
As the EMU0404 actually has around 113dBA SNR (see the measurements). Note that this is A-weighted so we need to reduce this a lot, around 10dB (estimate)to get linear noise.
So the AD conversion in the EMU0404 has around -103dB real noise. If you use the common formulas you find that for 17 bits you expects -104...
As the noise with signal in the FFT generally stays well above around -140dBFS, lets call it -138dBFS and run with it. The process gain for FFT pretty much bottoms out at around 33-36dB for normal FFT window sizes, so lets subtract 33dB from 138db and we get 105dB.
This is within 2dB of 103dB estimated above from the A-Weighted SNR, either way we are somewhere in the vicinity of 17 Bit's actual performance.
Incidentally, this "back of envelope math" also matches comparisons of using RMAA with known performance Audio Cards (I use a EMU1616M with RMAA Pro - the paid for version), you can find the loop-back results here (AD + DA that is):
http://audio.rightmark.org/test/EMU1616m.htm
Bottom line - Archimago's measurements may be adequate for CD Audio, they are meaningless for any other format. If he concludes "no difference" it simply means "no difference observable with inadequate test equipment and software".
QED.
Ciao T
Sometimes I'd like to be the water
sometimes shallow, sometimes wild.
Born high in the mountains,
even the seas would be mine.
(Translated from the song "Aus der ferne" by City)
"You are way off track. As you are already looking at an FFT you need to SUBTRACT (not add) process gain."How so? Did I misapply the paper I've linked or are you saying the paper is incorrect? The paper seems be pretty clearly saying the exact opposite of what you've stated when it states "The theoretical FFT noise floor is therefore 10log10(M/2) dB below the quantization noise floor due to the processing gain of the FFT." The go on to show an example where the process gain is added to the theoretical quantization noise floor limit of the ADC.
The math in this case for the 24 bit ADC in question goes like this: Theoretical quantization noise floor limit = 6.02(24)+1.76=146.24dB. Processing gain = 10log((2^24)/2)=69.24dB, putting the theoretical FFT noise floor for a 24bit ADC at near 215.48dB below full scale according to the paper.
Edits: 06/16/13
Hi,
OK, I misread you, my bad. I thought you inferred the -215dBFS from Archimagos posted FFT (noisefloor without signal).
So yes, if what Archimago used was a 24 Bit ADC we would expect a theoretical FFT noisefloor at a level much below -144dBFS.
As the process gain of the RMAA Software is unknown or at least undocumented, it can only be inferred by comparison with systems having known process gain (e.g. Audio Precision 2).
From that I make the process gain around 33dB, so a true 24 Bit ADC should show -144dB - 33dB = -177dB using RMAA. When I have the time I'll show what RMAA Shows for a 24 bit digital (SPDIF out to in) connection.
As we only see less than 140dB FFT noisefloor with signal we may conclude that Archimago's ADC (and it is the ADC in this case) is nearly 40dB short of 24 bit performance, or that it's performance is approximately 17 Bit.
So, no matter how we slice it and dice it, his test system is limited to around 17 Bit resolution and incapable making any determination about anything beyond.
By modern standards (my much better EMU1616M is the better part of a decade old!) this is pretty much rubbish.
Ciao T
Sometimes I'd like to be the water
sometimes shallow, sometimes wild.
Born high in the mountains,
even the seas would be mine.
(Translated from the song "Aus der ferne" by City)
stephan_g authoritatively declares the default RMAA process gain theoretically equal to 39dB in this linked thread. He seems fairly confident in his assertion for what it's worth. I wonder if M can be changed in RMAA somehow.
Edits: 06/17/13
Hi,
I never really looked.
If it is 39dB, then Archimago's test setup in not as...
as I thought it is, instead it is...
Ciao T
Sometimes I'd like to be the water
sometimes shallow, sometimes wild.
Born high in the mountains,
even the seas would be mine.
(Translated from the song "Aus der ferne" by City)
The process gain in analysis software is very easy to measure. Start with a file of white noise at a moderately high level. (You want it to be high enough that it is way above the noise floor of the numerical calculations themselves but low enough that no peaks ever clip.)
Now you average over a long time period. You want the noise file to be long enough that examining the spectrum at a point in the middle of the file doesn't change significantly as you move the point you have selected. (You will see end effects if you get too near the start or end of the file. If you aren't seeing these then you should continue to experiment until you do. Hitting an end of the file will change the window size, since it will cause the software to see lots of zero samples that aren't in the file or, worse, provoke a bug of some sort.)
You will now see the noise floor in the plotted graph. This will vary slightly as you move the sample point around in the middle of the file, particularly where you use a very high order FFT (e.g. 65K). This is normal and happens because each bin samples only a small amount of energy and the noise is (supposedly) random.
Your software will probably also be capable of computing some kind of an RMS average. You can look at this value. Here you want to average over a long interval so the results are stable. (Note that different software may give different RMS averages even though they both are doing an unweighted computation. That is because the idiots at the AES standardized on a zero level that corresponds to a sine wave, rather than a square wave. This is basic math and electrical engineering. So you may get results that differ by 3 dB.)
The process gain is just the difference you see between the single number calculated by the averaging program and the level of the noise floor read off the scale of the graphical spectrum plot.
If you try and calibrate the graphical scale by using test tones at known levels you may also observe funny effects. If a tone is in the dead center of an FFT bin it may plot at a different peak level than if the tone is straddling two bins and most of the energy is split between the two bins. In addition to changing with frequency this effect will also vary with the FFT size and window. This is because the size of the bins changes with FFT size and the shape of the bins changes with the window filter. (The window attempts to weigh points at the center greater than points at the edge to smooth out the effects of the last points measured. This is called the Gibbs effect.)
I generally do this kind of analysis with Soundforge. It is a good tool for playing around and understanding in a practical sense what is going on. Having a theoretical analysis of Fourier transforms, FFT algorithms, RMS averages, etc., is pretty close to useless when using tools for actual work. IMO it is much better, indeed essential, to have hands on experience with one's tools. If you feel the need to rely on some external authority for magic numbers such as "39 dB" you should realize that you need more hands on experience with the tools involved.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
Tony,
RMAA does not allow that.
It uses strictly it's own set of test tones and algorithms, it is "idiot proof" but as a result very limited in what you can and cannot do with it.
For example - you CANNOT feed it white noise and expect for it to do anything...
Ciao T
Sometimes I'd like to be the water
sometimes shallow, sometimes wild.
Born high in the mountains,
even the seas would be mine.
(Translated from the song "Aus der ferne" by City)
Post a Followup:
| FAQ |
Post a Message! |
Forgot Password? |
|
||||||||||||||
|
||||||||||||||
This post is made possible by the generous support of people like you and our sponsors:
Top of Page ]
[ Contact Us ]
[ Support/Wish List ]
[ Copyright © 2025, Audio Asylum, all rights reserved ]
