janneman said:
Jan,
the link shown at post #1
http://web.telecom.cz/macura/fmanalysis/freqdist.html
points to schematics of both amps:
http://web.telecom.cz/macura/classa_20w_en.html
and
http://web.telecom.cz/macura/pm_ab_er1.gif
As you can see, the classA has input stage with global NFB and classA push-pull output stage out of global NFB. The THD results for the output stage can be find as well:
http://www.pha.inecnet.cz/macura/next.htm
The error correction amp has global NFB. The output class AB output stage with error correction is placed inside the feedback loop of the AD844 CFB opamp.
Pavel
Graham Maynard said:
Graham,
the reason why I am trying this method is that I (and my visitors during blind tests) can clearly hear the difference after exchanging opamps like OPA627 and OPA134 in the preamp, for example, though distortion measured is less than -100dB.
For the FM measurement, please subtract distortion products amplitudes and original products peak amplitudes (about -20dB), you will get far less than 100dB difference.
Your question was about supersonics. Can we hear intermodulations of supersonic and audio band signals through tweeter? I do not know. But consider that similar mess is produced by CD players and preamps and it all goes to the power amp. It would not be filtered, as we do not speak about to high frequencies (< 50kHz).
Pavel
andy_c said:
Andy - thank you for your interest.
ad1) I will display it ASAP.
ad2) Yes, you are right. The test signal (input signal)
http://web.telecom.cz/macura/fmanalysis/mod_3.gif
was measured by loop-back method (card out - card in). It is directly fed to the input of the amps in both cases. Both amps are set for the same gain. The output of amps is divided by step attenuator before fed to the input of the sound card, to obtain similar level and to get into low own distortion range of the soundcard (-20 to -10 dBVrms). The own distortion of the soundcard can be seen here:
http://web.telecom.cz/macura/gen_kalib_6khz.gif
ad3) the conventional THD is about 2 - 3 times higher for the classA amp compared to error correction amp, but measured at lower frequency. I assume that the FM method represents at least THD and IMD together and thus is much more severe testing method for the amps. It is also limited not only to harmonic products like THD.
Pavel
Hi Pavel,
My take on the clearly audible differences between op-amps is that it is due to class-AB output stage and NFB/stabilisation circuitry. These affect 'dynamic' waveform capabilities even when the output stage loading is within stated limits, yet those effects are not measurable on steady sinewave THD.
I see you are attempting to visualise what we hear, but this still relates to steady sinewaves and not to the asymmetrical leading edges of music waveforms that must drive and charge circuit path capacitance via circuitry (op-amps) that are rendered inductive due to capacitive NFB/compensation/stabilisation circuitry. This is why output loading should be resistively limited, or external buffer transistors might still be beneficial.
I see that the 20W class-A has no NFB and that the cross coupled Q1+2 are driving capacitor linked CFPs with degenerative shunt connected 470pFs per half. The noise spectrum, as well as reproduction, might change if C2 is removed and C3/4 changed.
Cheers .......... Graham.
My take on the clearly audible differences between op-amps is that it is due to class-AB output stage and NFB/stabilisation circuitry. These affect 'dynamic' waveform capabilities even when the output stage loading is within stated limits, yet those effects are not measurable on steady sinewave THD.
I see you are attempting to visualise what we hear, but this still relates to steady sinewaves and not to the asymmetrical leading edges of music waveforms that must drive and charge circuit path capacitance via circuitry (op-amps) that are rendered inductive due to capacitive NFB/compensation/stabilisation circuitry. This is why output loading should be resistively limited, or external buffer transistors might still be beneficial.
I see that the 20W class-A has no NFB and that the cross coupled Q1+2 are driving capacitor linked CFPs with degenerative shunt connected 470pFs per half. The noise spectrum, as well as reproduction, might change if C2 is removed and C3/4 changed.
Cheers .......... Graham.
Improved null test
Pavel:
It seems to me you are both well equipped and motivated to perform a type of null test I proposed some time back. This test may eventually shed some light on the objectivist - subjectivist dilemma.
Conventional null tests are performed as far as I've seen in real time by hardware analogue means. This has several drawbacks among others the influence of linear distortions (that are distortions anyway) caused by phase/frequency response characteristics.
An alternative is to digitize and record input and output signals and post process them offline. A possible scheme could be:
1. Apply a simple test signal like the proverbial 1 kHz tone and digitize input and output.
2. Compute amplitude scaling and delay. This may be made directly dividing peak amplitudes and calculating the zero crossing difference, or better yet applying a brickwall filter algorithm to both datasets and finding the best fit parameters minimizing residual amplitude after subtracting the filtered, scaled and shifted datasets.
3. Repeat now with several representative segments of program material, using the calibrated parameters to obtain several difference datasets.
4. Apply this procedure to different amplifiers with known characteristics from the listening viewpoint.
5. For each dataset, make a waterfall time-frequency diagram. This is particularly important to unveil transient response signatures.
A further improvement should be to calibrate also the linear phase-frequency response and to include this model in the preprocessing of the output dataset before nulling, so as to remove its effect and keep only the nonlinearity related artifacts.
Hopefully the resulting diagrams may be a starting point at least to underscore differences. Whether they may be correlated to the listening experience is the challenge.
If you are willing to take the data collection part but have no time or do not want to take the data processing one, I am sure there will be volunteers here to help, including me within my possibilities.
Rodolfo
Pavel:
It seems to me you are both well equipped and motivated to perform a type of null test I proposed some time back. This test may eventually shed some light on the objectivist - subjectivist dilemma.
Conventional null tests are performed as far as I've seen in real time by hardware analogue means. This has several drawbacks among others the influence of linear distortions (that are distortions anyway) caused by phase/frequency response characteristics.
An alternative is to digitize and record input and output signals and post process them offline. A possible scheme could be:
1. Apply a simple test signal like the proverbial 1 kHz tone and digitize input and output.
2. Compute amplitude scaling and delay. This may be made directly dividing peak amplitudes and calculating the zero crossing difference, or better yet applying a brickwall filter algorithm to both datasets and finding the best fit parameters minimizing residual amplitude after subtracting the filtered, scaled and shifted datasets.
3. Repeat now with several representative segments of program material, using the calibrated parameters to obtain several difference datasets.
4. Apply this procedure to different amplifiers with known characteristics from the listening viewpoint.
5. For each dataset, make a waterfall time-frequency diagram. This is particularly important to unveil transient response signatures.
A further improvement should be to calibrate also the linear phase-frequency response and to include this model in the preprocessing of the output dataset before nulling, so as to remove its effect and keep only the nonlinearity related artifacts.
Hopefully the resulting diagrams may be a starting point at least to underscore differences. Whether they may be correlated to the listening experience is the challenge.
If you are willing to take the data collection part but have no time or do not want to take the data processing one, I am sure there will be volunteers here to help, including me within my possibilities.
Rodolfo
Pavel,
Your post is a good contribution. The question of what measurements would better correspond to sound quality preferences has been a hot topic for decades, and some of us have not been too pleased with the tests commonly used. Your results look promising. I intend to try it myself.
I do offer a caution, however (though likely a non-issue in this case). The method of feeding the A/D portion of the system through a resistive divider will present a rather high source impedance to the input analog stage. Some such stages (op-amp based non-inverting stages in particular) may have a distortion characteristic that depends on driving impedance. To ensure a fair comparison, the effective driving impedance should be made equal (in addition to achieving the appropriate signal amplitude).
I am hoping to see additional posts of such measurements by others for some of the common audio circuits (preamps, chip amps). I may do my own testing in about six months or so.
Dale
Your post is a good contribution. The question of what measurements would better correspond to sound quality preferences has been a hot topic for decades, and some of us have not been too pleased with the tests commonly used. Your results look promising. I intend to try it myself.
I do offer a caution, however (though likely a non-issue in this case). The method of feeding the A/D portion of the system through a resistive divider will present a rather high source impedance to the input analog stage. Some such stages (op-amp based non-inverting stages in particular) may have a distortion characteristic that depends on driving impedance. To ensure a fair comparison, the effective driving impedance should be made equal (in addition to achieving the appropriate signal amplitude).
I am hoping to see additional posts of such measurements by others for some of the common audio circuits (preamps, chip amps). I may do my own testing in about six months or so.
Dale
Oddly enough, the school of thought that argues even the smallest measurable noise or distortion is measurable appears sometimes allied with the school of thought that there are magic gadgets that have audible properties that transcend measurement.
In any case, to strive toward ever lower noise and distortion figures is a pursuit that brings joy and satisfaction to the designer. So long as other considerations such as such as reliability and saftey are not compromised, it does little harm.
Meanwhile the "next frontier" would seem to me to be reducing speaker distortion.
goudey said:
Dale,
you are right, but I use the divider behind the output of the power amp (power amp input is driven directly from low impedance), just to reduce the input signal into the sound card. As the divider was the same for both amps compared, I do not think there was a problem.
Pavel
Another veiwpoint
Seems to me that what PMA's evidence is suggesting is that it's not so much gross distortion (IM, THD, Phase)as it is something else. What Maynard says makes sense. If so this doesn't put the standard battery of brute, outdated tests in a favorable light. Giving Mr. Maynard the benefit of the doubt, perhaps it would behoove audio enthusiasts to develop a test with the previous stated criteria. <what Ingrast said>
The only issue I have with you post is that a 1kHz is worth less than a pile of ....stinky doo
, when looking at amplifier subtleties. You need a complex waveform. a la music. Otherwise, what's the point? Seriously!
I was thinking on this for a few minutes and it occurred to me, wouldn't it be possible to develop a test, that takes into Maynard's posting, and use a computer to measure the differences between the two signals?
ex. Computer measures the waveform of the smaller input signal's "Asymmetrical leading edges of music waveforms" with the greater amplified output.
In this way, we can "visualize what we hear" graphically via datasets and graphically through a spreadsheet perhaps?
One could then isolate the offending amplification stage and correct and remeasure, a la Douglas Self.
(Self's amp has been reported to be NOT perfect. Therefore I would say that his research methodology is not in error but that he has NOT discovered or ackknowledged the distortion mechanisms that are causing people to label his amp as being less than perfect.)
The logical conclusion for me, assuming that there is an audible difference, is that if you hear it, then you can measure it (the opposite is not necessarily true) . If you hear a difference and cannot measure it. You aren't looking at the right errors. Giving that the dynamic range of human hearing is about 120dB (ie. atomic pressure to airplane jets) I would be inclined to say that -100dB distortion is on the edge of human perceivability and any other distortions you might hear are from other mechanisms. Although our animal companions might take issue! I think you'll agree that standard distortion tests only show a limited amount of distortion causes/effects.
David
P.S. Psychoacoustics not withstanding
P.S.S. The only major, insurmontable? problem I see is having a perfect ADC into the analysis program as to rule out it as a distortion contributor... is it possible?
Seems to me that what PMA's evidence is suggesting is that it's not so much gross distortion (IM, THD, Phase)as it is something else. What Maynard says makes sense. If so this doesn't put the standard battery of brute, outdated tests in a favorable light. Giving Mr. Maynard the benefit of the doubt, perhaps it would behoove audio enthusiasts to develop a test with the previous stated criteria. <what Ingrast said>
The only issue I have with you post is that a 1kHz is worth less than a pile of ....stinky doo
, when looking at amplifier subtleties. You need a complex waveform. a la music. Otherwise, what's the point? Seriously!I was thinking on this for a few minutes and it occurred to me, wouldn't it be possible to develop a test, that takes into Maynard's posting, and use a computer to measure the differences between the two signals?
ex. Computer measures the waveform of the smaller input signal's "Asymmetrical leading edges of music waveforms" with the greater amplified output.
In this way, we can "visualize what we hear" graphically via datasets and graphically through a spreadsheet perhaps?
One could then isolate the offending amplification stage and correct and remeasure, a la Douglas Self.
(Self's amp has been reported to be NOT perfect. Therefore I would say that his research methodology is not in error but that he has NOT discovered or ackknowledged the distortion mechanisms that are causing people to label his amp as being less than perfect.)
The logical conclusion for me, assuming that there is an audible difference, is that if you hear it, then you can measure it (the opposite is not necessarily true) . If you hear a difference and cannot measure it. You aren't looking at the right errors. Giving that the dynamic range of human hearing is about 120dB (ie. atomic pressure to airplane jets) I would be inclined to say that -100dB distortion is on the edge of human perceivability and any other distortions you might hear are from other mechanisms. Although our animal companions might take issue!
I think you'll agree that standard distortion tests only show a limited amount of distortion causes/effects.David
P.S. Psychoacoustics not withstanding
P.S.S. The only major, insurmontable? problem I see is having a perfect ADC into the analysis program as to rule out it as a distortion contributor... is it possible?
Re: Another veiwpoint
Absolutely, what I propossed the 1kHz tone for was only for calibration of the null process setup.
Undoubtedly the real test is with actual program source. Perhaps I was not sufficiently clear in my previous post.
The final performance figure of merit if you want, is for example a waterfall 3D diagram (time/frequency/amplitude) of the null result dataset for selected passages (those audiophiles may point to to expose a certain amplifier strength or weakness) and see how it correlates if at all.
Rodolfo
EternaLightWith said:.... it would behoove audio enthusiasts to develop a test with the previous stated criteria. <what Ingrast said>
The only issue I have with you post is that a 1kHz is worth less than a pile of ....stinky doo
Absolutely, what I propossed the 1kHz tone for was only for calibration of the null process setup.
Undoubtedly the real test is with actual program source. Perhaps I was not sufficiently clear in my previous post.
The final performance figure of merit if you want, is for example a waterfall 3D diagram (time/frequency/amplitude) of the null result dataset for selected passages (those audiophiles may point to to expose a certain amplifier strength or weakness) and see how it correlates if at all.
Rodolfo
Hi David,
In my EW article I did suggest a direct and 'live' differencing comparison of loudspeaker terminal voltage against a digitally (group) delayed 24/196 copy of amplifier input, but the X-Y method also gives a 'live' display, with an insulated scope of course.
When X-Y tested with 10kHz sinewave the Self type of Blameless amp produces a smooth but awfully large 'Y' error which suggests that a stereo image will be unclear.
Hi Rodolpho,
Waterfall ? This sounds a good idea but is it acheivable ?
Given that most power amplifiers distort more when the load is a real loudspeaker and not a resistor, also that not all first or second or later cycles induce the same degree of interface distortion, how can the waterfall cover this from amplifier input ?
Cheers .......... Graham.
In my EW article I did suggest a direct and 'live' differencing comparison of loudspeaker terminal voltage against a digitally (group) delayed 24/196 copy of amplifier input, but the X-Y method also gives a 'live' display, with an insulated scope of course.
When X-Y tested with 10kHz sinewave the Self type of Blameless amp produces a smooth but awfully large 'Y' error which suggests that a stereo image will be unclear.
Hi Rodolpho,
Waterfall ? This sounds a good idea but is it acheivable ?
Given that most power amplifiers distort more when the load is a real loudspeaker and not a resistor, also that not all first or second or later cycles induce the same degree of interface distortion, how can the waterfall cover this from amplifier input ?
Cheers .......... Graham.
Graham Maynard said:
Graham,
A "waterfall" diagram is a 3D surface plot where for example the x and y axis represent time and frequency, while the z or vertical axis represents amplitude. Here you will find in this tweeter spec sheet, a typical waterfall type of diagram in this case for the spectral decay of the driver after exited by an impulse (bottom of page 2).
In the case of a null test output dataset, the diagram should expose the running spectral composition of the *difference* between the amplifier input and output, that is, the nature of the stuff added or substracted from the original signal.
Certainly the ultimate null test should comprise the speakers themselves so you need a lab quality microphone and anechoic room ...
Things require a begining though.
Rodolfo
Regarding correlation between measurements and perception, I have the following link to contribute. Sorry if this has already been posted.
http://www.gedlee.com/distortion_perception.htm
http://www.gedlee.com/distortion_perception.htm
PMA,
You were a participant in the “What's your reasoning?" and not "What's your belief?" thread, did you miss the Czerwinski ref?
http://www.diyaudio.com/forums/showthread.php?postid=492032#post492032
also your fm mod source – loop-back pict looks nothing like my simulated fm modulation spectrum
http://www.diyaudio.com/forums/showthread.php?postid=485573#post485573
we should expect a 10KHz central peak with +/- n*2KHz decreasing sidebands from your test signal
if you had read the Czerwinski article you should understand about IMD masking in test signals with integer multiple frequency differences – like the fm modulation signal
if you can’t access JAES articles then a web based explanation and test signal recommendation can be found on jon risch’s site
http://www.geocities.com/jonrisch/page10.htm
for more, try googling multitone, multisine + distortion or nonlinear system identification
photobucket seems to be down at the writing of this post so I’ll attach the fm mod pict from my early “What's your reasoning?" and not "What's your belief?" post #80
You were a participant in the “What's your reasoning?" and not "What's your belief?" thread, did you miss the Czerwinski ref?
http://www.diyaudio.com/forums/showthread.php?postid=492032#post492032
also your fm mod source – loop-back pict looks nothing like my simulated fm modulation spectrum
http://www.diyaudio.com/forums/showthread.php?postid=485573#post485573
we should expect a 10KHz central peak with +/- n*2KHz decreasing sidebands from your test signal
if you had read the Czerwinski article you should understand about IMD masking in test signals with integer multiple frequency differences – like the fm modulation signal
if you can’t access JAES articles then a web based explanation and test signal recommendation can be found on jon risch’s site
http://www.geocities.com/jonrisch/page10.htm
for more, try googling multitone, multisine + distortion or nonlinear system identification
photobucket seems to be down at the writing of this post so I’ll attach the fm mod pict from my early “What's your reasoning?" and not "What's your belief?" post #80
Attachments
jck,
When I simulate the FM modulated signal as described (at 96khz sample rate) I get essentially what was graphed.
The standard FM modulation formula is:
t = [0:M]/fs;
sin( 2*pi*f_center*t+ f_deviation / f_mod_frequency * sin(f_mod_frequency*2*pi*t));
The "weighting function" I used was (in matlab-like terms):
x = [0:M];
w = ones(1,M+1) - 4/3 * cos(2*pi/(M+2)*1*(x+1)) + 1/3 * cos(2*pi/(M+2)*2*(x+1));
A "weighting function" does need to be applied to control "spectral leakage". It would be nice to "standardize" the waveform generation and processing to allow for apples-to-apples comparisons for tests performed by different people.
Pavel,
Some amplifiers are more particular than others regarding the driving impedance of the source. An amplifier may perform well with a low source impedance, and poorly with a high source impedance. We won't know what is the case unless a suitable test is performed.
Dale
When I simulate the FM modulated signal as described (at 96khz sample rate) I get essentially what was graphed.
The standard FM modulation formula is:
t = [0:M]/fs;
sin( 2*pi*f_center*t+ f_deviation / f_mod_frequency * sin(f_mod_frequency*2*pi*t));
The "weighting function" I used was (in matlab-like terms):
x = [0:M];
w = ones(1,M+1) - 4/3 * cos(2*pi/(M+2)*1*(x+1)) + 1/3 * cos(2*pi/(M+2)*2*(x+1));
A "weighting function" does need to be applied to control "spectral leakage". It would be nice to "standardize" the waveform generation and processing to allow for apples-to-apples comparisons for tests performed by different people.
Pavel,
Some amplifiers are more particular than others regarding the driving impedance of the source. An amplifier may perform well with a low source impedance, and poorly with a high source impedance. We won't know what is the case unless a suitable test is performed.
Dale
goudey
looks like i transposed 2K/200
but my point is that the "ordered complexity" of PMA's fm test tone, with many n*fmod sidebands will have large, distortion obscuring signals at many of the IMD product frequencies - and many more IMD product frequencies will be the summation of multiple orders of IMD from many of the (equally spaced) fm sidebands
Czerwinski is really the place to start if you hope to understand, much less advance distortion testing
Risch 's "split phi" test tone multisines keep the distortion overlaps down to a level that eyeball inspection can at least draw something from
also to uncover possible low frequency intermod effects you need widely seperated stimuli frequencies as well
looks like i transposed 2K/200
but my point is that the "ordered complexity" of PMA's fm test tone, with many n*fmod sidebands will have large, distortion obscuring signals at many of the IMD product frequencies - and many more IMD product frequencies will be the summation of multiple orders of IMD from many of the (equally spaced) fm sidebands
Czerwinski is really the place to start if you hope to understand, much less advance distortion testing
Risch 's "split phi" test tone multisines keep the distortion overlaps down to a level that eyeball inspection can at least draw something from
also to uncover possible low frequency intermod effects you need widely seperated stimuli frequencies as well
Hi Rodolpho,
I am well aware of waterfalls, but I see a time related polarity element being necessary too, such that a single 3D plot cannot reveal all of an amplifier's activity, especially when propagation delayed and NFB loop controlled rising responses are different to falling responses as they react to constant dynamic alteration of load angle.
Different + and - responses follow from specific and momentary waveform/loudspeaker responses and I wonder how the waterfall might integrate these.
Maybe a dual waterfall display with centre frequency zero, neg to the left and pos to the right.
Cheers ........... Graham.
I am well aware of waterfalls, but I see a time related polarity element being necessary too, such that a single 3D plot cannot reveal all of an amplifier's activity, especially when propagation delayed and NFB loop controlled rising responses are different to falling responses as they react to constant dynamic alteration of load angle.
Different + and - responses follow from specific and momentary waveform/loudspeaker responses and I wonder how the waterfall might integrate these.
Maybe a dual waterfall display with centre frequency zero, neg to the left and pos to the right.
Cheers ........... Graham.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.