Happy New Year to all, and remember this Slovak phrase:
" How many languages you know, so many times you are a human".😉
It is nothing wrong (IMO) with more than 2000 years old latin phrase, if it is stil valid..
" How many languages you know, so many times you are a human".😉
It is nothing wrong (IMO) with more than 2000 years old latin phrase, if it is stil valid..
A proud owner of a sound card toy (at least when it comes to instrumentation) was challenging me to show the noise floor of one of these 16 bit spectrum analyzer oldies. Here's what one of these boat anchors can do (attached). As you see, the noise floor is around -170dB 🙂.
Now, of course, some explanations are required. Sound card based spectrum analysis software do what is called "rms averaging". In this mode, the software takes the result of the last measurement, adds it to the previous measurements (point by point) and then takes the rms sum of these to compute the average. This is wrong to start with (since the right way to do it is to multiply the spectrum with its conjugate) but anyway, call this pure scalar method "acceptable".
However, whatever way it's done, rms averaging (and sorry for shouting) DOES NOT ELIMINATE NOISE, it only produces an approximation of the actual noise level. Increasing the number of rms averages provides simply a better statistical approximation of the noise but (sorry again) IT WILL NOT ACTUALLY LOWER THE NOISE. With rms averaging, the spectral components also include noise. For small signals, this noise can add significantly to the spectral components magnitudes. You CAN'T eliminate this error source by using rms averaging.
If it is to reduce noise, vector averaging is required. In vector averaging mode, the $500 boot anchor averages complex values in the frequency domain. This lowers the noise, since the real and imaginary components of the random signals (as noise is) are not in phase, and therefore cancel each other - increasingly so with each average. Frequency components that are periodic do not cancel, and therefore do not diminish with successive averaging. Of course, the noise added to each spectral component is also averaged/reduced.
One requirement for a successful vector averaging is the ability to provide a trigger signal. These oldies have of course all the triggering features known to mankind, including triggering on its own signal source. Without triggering, there is no guarantee that the spectral components are always in phase during averaging, therefore they will also cancel. Now, if anybody is still wondering why a sound card won't do vector averaging, it should be pretty clear: I still have to see such consumer level hardware having any form of trigger I/O 🙂.
Bottom line, this 16 bit HP 35665A boat anchor, a senior of mine, has a floor noise of under -170dB after only 6400 averaging (about 10 minutes). The maximum number of averages is 65535, or ten times, and as the noise power goes down with the root square of the number of averaging, you do the math for the minimum noise floor achievable, and compare with your best soundcard toy.
Now, of course, some explanations are required. Sound card based spectrum analysis software do what is called "rms averaging". In this mode, the software takes the result of the last measurement, adds it to the previous measurements (point by point) and then takes the rms sum of these to compute the average. This is wrong to start with (since the right way to do it is to multiply the spectrum with its conjugate) but anyway, call this pure scalar method "acceptable".
However, whatever way it's done, rms averaging (and sorry for shouting) DOES NOT ELIMINATE NOISE, it only produces an approximation of the actual noise level. Increasing the number of rms averages provides simply a better statistical approximation of the noise but (sorry again) IT WILL NOT ACTUALLY LOWER THE NOISE. With rms averaging, the spectral components also include noise. For small signals, this noise can add significantly to the spectral components magnitudes. You CAN'T eliminate this error source by using rms averaging.
If it is to reduce noise, vector averaging is required. In vector averaging mode, the $500 boot anchor averages complex values in the frequency domain. This lowers the noise, since the real and imaginary components of the random signals (as noise is) are not in phase, and therefore cancel each other - increasingly so with each average. Frequency components that are periodic do not cancel, and therefore do not diminish with successive averaging. Of course, the noise added to each spectral component is also averaged/reduced.
One requirement for a successful vector averaging is the ability to provide a trigger signal. These oldies have of course all the triggering features known to mankind, including triggering on its own signal source. Without triggering, there is no guarantee that the spectral components are always in phase during averaging, therefore they will also cancel. Now, if anybody is still wondering why a sound card won't do vector averaging, it should be pretty clear: I still have to see such consumer level hardware having any form of trigger I/O 🙂.
Bottom line, this 16 bit HP 35665A boat anchor, a senior of mine, has a floor noise of under -170dB after only 6400 averaging (about 10 minutes). The maximum number of averages is 65535, or ten times, and as the noise power goes down with the root square of the number of averaging, you do the math for the minimum noise floor achievable, and compare with your best soundcard toy.
Attachments
Last edited:
dedicated hardware is better, more convenient worth the money to pros, their employers paying by the hour
for advanced hobbyists however $k hardware taking up shelf space that get turned on few 10s of hours a year may want to go cheap, deal with limitations of soundcards
and they aren't so limited if you use good measurement technique, adapt to their features, limits
latency may not be guaranteed but is fixed within a single capture, is fixed between channels
use one of the stereo ADC channels to capture the DAC output test signal and you do have a good phase ref
simply going long in the dataset gets you noise floor improvement from it being spread over many finer bins - I've used 100 second captures, fft in SciLab (a free MatLab workalike) - did the math with the full complex fft arrays
averaging randomly started acquisitions has a possible advantage of sometimes reducing spurs, interference lines - but for the most part I just look at the loopback and use structured enough signal that I can calculate the harmonics and IMD products characteristic of locally Lipschitz nonlinearities, avoid spurs like power line frequency harmonics, high frequency lighting switching freqs if distortion from ppm nonlinearities is what I want to look at with audibly pointlessly ridiculous resolution
for advanced hobbyists however $k hardware taking up shelf space that get turned on few 10s of hours a year may want to go cheap, deal with limitations of soundcards
and they aren't so limited if you use good measurement technique, adapt to their features, limits
latency may not be guaranteed but is fixed within a single capture, is fixed between channels
use one of the stereo ADC channels to capture the DAC output test signal and you do have a good phase ref
simply going long in the dataset gets you noise floor improvement from it being spread over many finer bins - I've used 100 second captures, fft in SciLab (a free MatLab workalike) - did the math with the full complex fft arrays
averaging randomly started acquisitions has a possible advantage of sometimes reducing spurs, interference lines - but for the most part I just look at the loopback and use structured enough signal that I can calculate the harmonics and IMD products characteristic of locally Lipschitz nonlinearities, avoid spurs like power line frequency harmonics, high frequency lighting switching freqs if distortion from ppm nonlinearities is what I want to look at with audibly pointlessly ridiculous resolution
Last edited:
Attached is the same 16 bit HP 35665A boat anchor analyzing the output of a Tektronix SG505 low distortion oscillator, set at 10KHz and 1Vrms. The fundamental is notched, the spectrum is measured using vector averaging, trigger is on Channel 1. The output is mathematically corrected for the notch filter gain/attenuation, measured separately, and stored as a constant K1. Only a few averages this time, enough to make my point, not much time to spend on this.
3rd harmonic is dominating the distortions, at about -110dB, 2nd is under -130dB, everything else is in the noise floor. A Rohde UPV distortion analyzer measured, on the same signal, -110dB THD10 in a 100KHz bandwidth. Obviously, with a good notch filter (like in this example) and long averaging, measuring distortions down to -160dB is quite possible. Mr. Cordell may like this kind of setup for his study of distortions in metal film resistors. In your face, sound cards 😉.
I rest my case.
3rd harmonic is dominating the distortions, at about -110dB, 2nd is under -130dB, everything else is in the noise floor. A Rohde UPV distortion analyzer measured, on the same signal, -110dB THD10 in a 100KHz bandwidth. Obviously, with a good notch filter (like in this example) and long averaging, measuring distortions down to -160dB is quite possible. Mr. Cordell may like this kind of setup for his study of distortions in metal film resistors. In your face, sound cards 😉.
I rest my case.
Attachments
JCX, a HP 35665A can be purchased on your side of the pond for less the cost of a Linx L22 sound card*). Anyway, my beef is with those claiming that sound cards are the best thing since sliced bread, and that they exceed the performance of any 30 years old 16 bit analyzer.
I've already mentioned that sound cards are very good for the DIY enthusiasts, in particular if an interface is built. BTW, I measured Peter Millet interface that I mentioned in this thread and, as expected, THD at >10 KHz is a disappointing 10-15 ppm (oh sorry, I forgot, Mr. Ueber-moderator doesn't care about these levels). I'll show the results when I'll have the time to process them.
P.S. Yes, this 35665A is indeed good as a boat anchor. It weights about 40 kilos, 46dB (re: 1Kg) over a L22. In performance per kilo, sound cards are a clear winner.
*) Example: Agilent 35665A Dynamic Signal Analyzer 2 Channel DC to 102 4 kHz Opt 1D4 | eBay
Last edited:
I paid $130 for my ESI Juli@ - Lynx are pricey for the performance - I just squint and the Juli@ is good enough for audio
cheap vector impedance for component/layout parasitics at >10x loop gain intercept - few hundred MHz would be nice though
cheap vector impedance for component/layout parasitics at >10x loop gain intercept - few hundred MHz would be nice though
Last edited:
Sure video averaging won't reduce the noise floor, and soundcards can't do synchronous averaging. But you can run a very large FFT size, and that does reduce the noise floor in a similar way to vector averaging.
Irrespective of what DSP techniques you use, the distortion measuring performance is ultimately limited by the linearity of the ADC and analog front end.
Irrespective of what DSP techniques you use, the distortion measuring performance is ultimately limited by the linearity of the ADC and analog front end.
True. Did I mention that, chances are, a 16 bit SAR ADC has less differential and integral nonlinearities compared to a 24bit delta-sigma audio grade ADC?
I wouldn't be surprised if 16 HP bits were worth more than 24 consumer ones, but I wouldn't just assume this either.
I believe they can, at least using ARTA software in single channel mode. I've never actually had a need to do that, so don't have examples to show; in audio, you generally don't want to average out the noise floor of the DUT so RMS averaging is more appropriate.
flagship delta-sigma audio ADC/DAC do differential linearity at SOTA for single chip monolithic solutions today - usually have relative smooth curved transfer error
R2R, successive approximation converters suffer from major bit tolerance mismatch giving the worst possible characteristic for audio measurement
but looking at your full amplitude test signal and the ppm distortion products together is asking a bit much - notch filter, knock down the test tones or bandpass just the distortion product freq
R2R, successive approximation converters suffer from major bit tolerance mismatch giving the worst possible characteristic for audio measurement
but looking at your full amplitude test signal and the ppm distortion products together is asking a bit much - notch filter, knock down the test tones or bandpass just the distortion product freq
😛.
Waly , You are candidate for next year Nobel price , 170dB resolution with 16bit sampling, it is outside known physics laws..
Try to show real dynamic range, measurement with e.g. 1kHz signal at 0dBV and what will be the noise background than🙄 Do not "fidle" with notch filter, external generators and other "aids" Show measurmement for 16bit analyzer, "all in one"....
Attached is measurment i did now, made in 10second, only souncard needed. Fast, convenient, reliable...
And I am curious how you will do sweep measurements, CCIF, DIM, multitone..
Attachments
Last edited:
With all due respect, you have no idea what you are talking about. I could recommend a few authoritative references on this matter, but I'm afraid it would be a waste of time as long as you lack the very basic knowledge of digital signal processing. Stay in awe with your favourite toy, ignorance is bliss...
With all respect ,too, do not isolate one part of my post from other..Show dynamic range of such measurement without notch filter... And here were also some other questions, or ?? Sweeeps e.t.c?? I am still waiting..
You should know that I used such vintage analyzers more than 15 years, it is not on level of todays needs..
I attach another measurement with the same card, other software package with vector averaging used, simple card loopback in analog domain (D/A, A/D duplex), no other aids needed..
Quality sound card is very valuable tool (not toy..), if one know how to use it properly and get the maximum..
You should know that I used such vintage analyzers more than 15 years, it is not on level of todays needs..
I attach another measurement with the same card, other software package with vector averaging used, simple card loopback in analog domain (D/A, A/D duplex), no other aids needed..
Quality sound card is very valuable tool (not toy..), if one know how to use it properly and get the maximum..
Attachments
Last edited:
Wally you're a smart guy you certainly know that the time domain and frequency domain representation of the signal contain the same information and that averaging noise in the time domain is an exact equivalent of vector averaging the FFT's and taking the inverse transform. Since random noise has a zero mean and does not correlate on any time scale synchronous averaging is not necessary to generate a plot of noise averaged to, in the limit, its mean value at every point. Or in other words zero (- inf dB).
As to BV's point if you want to add resolving distortion components on a signal generated at the 16bit level you will have to consider the dither technique and its implication in the achievable resolution. Certainly an undithered 16bit sin wave puts one at a serious disadvantage. I had a discussion with Bob Adams about the concept of synchronous dither that would fall out of the picture but we didn't get too far.
I also wish folks would stop talking about "sound card software" A sound card is A/D's, D/A's and a device driver. Modules/plug-in's are available for NI, Matlab, Octave, Python, etc. for streaming the data as if it was from any instrument. This is one place where the arbitrary N FFT comes in handy, at 96k exactly 1k Hz, for example, is possible so a stream of data can be averaged in some block that is an exact multiple of 96 samples and then take the FFT. Since the phase is arbitrary but the same for every block the averaging is synchronous.
I have used sound cards to generate proof of concept data to BIG customers in applications that have nothing to do with audio. There are some external devices with balanced XLR inputs and the AKM chipset has a pin you can lift to DC couple the A/D. This is a mod not for the faint of heart because the pins are tiny and failure leaves you with a toaster.
But in the end the distortion magnifier or twin-t techniques work quite well, as well as jcx's separation by filtering technique for IMD. We have used this for xDSL testing using band separation filters from Allen Avionics.
BTW your attitude still needs some work...
Last edited:
- Status
- Not open for further replies.