Geddes on distortion measurements

[Gordy]

It's not an arguement with concrete data, but some typical experiments make the essence of the statement clear.

( ... snip... )

As I say we can argue the percentages (but I won't) but there is little doubt (in my mind at least) that the vast majority of the audio reproduction problem is the loudspeaker. Experience says this, the data says this and most other researchers (like Floyd Toole) say this.

Thank you. Interesting. I think you may have over-simplified, but I understand your sentiment. I'll go read-up some more...

 

[gedlee]

I think you may have over-simplified...

The thing that I most dislike about these posts is that one virtually has no choice but to oversimplify things. I spend a few minutes on a post. It took me a year to write my book - full time. How can you NOT oversimplify?

 

[MBK]

Attached are some examples of the output. The signals are scalled upand right to offset them for better observation. The only one on the correct scale is the lowest one - red.

Look for no change in harmonic structure of the waveform as the signal level drops. These should not be any very high harmonics.

The Sanyo is clearly a disaster, the Sony is not very good as the harmonics are higher order and are pretty predominate at mid levels. The Pioneer(my amp) shows a very clean response all the way down in signal level.

Earl,

very revealing, never seen such a good visualization of low level distortion. Actually, I don't think I *ever* saw a quantification of low level distortion that was truly convincing.

People who hear differences between amps that measure "identically" (on simple tests) are often derided as subjectivists - and to be fair it's just really hard to pin anything down without data. This kind of test is really the missing link.

I'd really like to use this technique. I hope we can find a resolution to the IP issue.

Usually an attached note such as "Copyright ... only private use ... use with credit to... " is legally enough, and that's how the software guys do it, BUT this does not make it impossible to rip it off. And then the only option left is to sue if you want your rights respected.

Other option of course is to PM the WAV only to interested people, collect their names and addresses and make them send you an aceptance note of terms and conditions for use (as in software the usual "I accept" button) .

 

[AJinFLA]

People who hear differences between amps that measure "identically" (on simple tests) are often derided as subjectivists

I've never seen anyone derided when science enters the picture and controls are used when listening to different amps. Have you?
I've seen plenty of sighted, uncontrolled, unscientific listening comparisons made, followed by derision.
I bet lots of folks "heard" how poorly Earls Pioneer sounded at RMAF...the second they saw it.
Maybe Earl needs to file a patent fast. Looks like an interesting test indeed.

cheers,

AJ

 

[gedlee]

Other option of course is to PM the WAV only to interested people, collect their names and addresses and make them send you an aceptance note of terms and conditions for use (as in software the usual "I accept" button) .

I think that I'd like to track who I give the IP to. That, I think, makes the most sense.

The software runs in MathCad. I am not well versed in MathCads free "run time only" software. Is anybody familiar with that?

I think that if I require a direct E-mail with a statement agreeing to not use the software and concepts without noting the orignator that should suffice. "Egeddes@gedlee.com"

If you want a copy of the wav file and the MathCad analysis, then E-mail me directly with a statement - in your own words - that you will not distribute it without my permission. I'll send you a copy of both. If you don't have MathCad and can't run it, I don't know what to do since the calculations are rather extenisve and can't really be done in Excel.

 

[gedlee]

I bet lots of folks "heard" how poorly Earls Pioneer sounded at RMAF...the second they saw it.

I couldn't have said it better myself!! Thanks

 

[Brett]

Earl,

very revealing, never seen such a good visualization of low level distortion. Actually, I don't think I *ever* saw a quantification of low level distortion that was truly convincing.

People who hear differences between amps that measure "identically" (on simple tests) are often derided as subjectivists - and to be fair it's just really hard to pin anything down without data. This kind of test is really the missing link.

I'd really like to use this technique. I hope we can find a resolution to the IP issue.

Usually an attached note such as "Copyright ... only private use ... use with credit to... " is legally enough, and that's how the software guys do it, BUT this does not make it impossible to rip it off. And then the only option left is to sue if you want your rights respected.

Other option of course is to PM the WAV only to interested people, collect their names and addresses and make them send you an aceptance note of terms and conditions for use (as in software the usual "I accept" button) .

I'm in complete agreement with Martin. That is the best, and only that I ever recall, clear example of increasing distorion with decreasing level. I'd often wondered about using an averaging technique as is common in RF, but don't have the skills to code it.

If something does become available, I would happily sign off on a user agreement, and give credit where it's due.

I've learnt a lot form many of your posts Earl. Cheers.

 

[gaetan8888]

The amp is a Pioneer VSX-V912 - a receiver. The point is that it was on sale at Costco for $150.00. I bought several of them for home theater use. I used my test to measure the amps and they were quite good actually. Especially for chip amps. I was measuring a lot of chip amps (a survey of capability) and most were pretty bad. As a chip amp this unit deffinately stands out. It compared quite favorably to a very well engineered discrete amp that I also use.

I also tested several other receivers and they were almost universally bad.

Hello

Did you check wich chip amp are inside your Pioneer VSX-V912 ?

Thank

Gaetan

 

[gedlee]

Did you check wich chip amp are inside your Pioneer VSX-V912 ?

It says Pioneer on it. It appears to be their design or done for them.

 

[MBK]

I think the user agreement solution is perfectly reasonable, I'd be happy to send one.

On the practical side though, do I get this straight:

1- does one need just the .wav file and a typical soundcard based FFT package, or

2- does one also need to run MathCad to read/write/recalculate the data for the test?

2- would be tougher, MathCad runs at USD1000 these days and I couldn't even find any current stripped down MathCad player on their site. They do give you a 30 day eval demo though.

Also, how do you secure synchronization of the in and outputs to each other and to the A/D and D/A (since this is the crux to get the noise floor down)?

 

[gedlee]

I think the user agreement solution is perfectly reasonable, I'd be happy to send one.

1- does one need just the .wav file and a typical soundcard based FFT package, or

You only need to record the output with the sound card. The FFT's run in the analysis program.

2- does one also need to run MathCad to read/write/recalculate the data for the test?

At the present you do need MathCad. There used to be a free viewer for Mathacd which would allow you to run a file but not edit it. This would be fine, even prefered.

If MathCad is not an option, I am well versed in reading and writing wav files in VB and FORTRAN so I can do an independent piece of software, but my preference is for MathCad since its already done.

Also, how do you secure synchronization of the in and outputs to each other and to the A/D and D/A (since this is the crux to get the noise floor down)?

When I went back and looked over what I had done - about a year ago - I find that I had incorrectly stated the details. As I looked over the calculations, I'm not precisely sure what I actually ended up doing or why. (They say that memory is the second thing to go with age! God, I hope its the first!!! I just discovered Bangkok!)

I will repost when I get the details sorted out.

In the mean time please look for a MathCad viewer as I know that they existed at one point.

 

[jcx]

Cheap soundcard, Free Software

I have done some testing using the US$130 ESI Juli@ at 24/192

averaging is equivalent to simply using long records to lower the fft noise floor

dithering is unecessary with good performance 24 bit ADC/DAC so cheaply available today

I have created minute long tests with 6 second steps at a single frequency and performed "milliion" point ffts (2^20 actually)

LtSpice is handy for creating .wav at high resolution and with BV source the function can be highly tailored - step/sweep frequency, phase, amplitude

I then use SourceForge Audacity to play/record the test wavs

http://audacity.sourceforge.net/

LtSpice is somewhat limited for analysis but it can do large ffts with good resolution Blackman windowing

I've used the free SciLab software, a MatLab "work alike" to process the measurement wav (waveform creation obviously doable too)

http://www.scilab.org/

SciLab has "extended" wavread/write function on their site for download to deal with 24 or 32 bit wav (default sound/wav functions are 16 bit) - a useful trick is to save results as 32 bit wav in Audacity - 24 bit xwavread is slow and memory wasting in SciLab

 

[MBK]

OK, googled and found MathCad Explorer, version 8 is the last available.

BTW a 79 yr old friend once told me, you know you're old if you forget to close your zipper after visiting the bathroom.

You know you're really old when you forget to open it.

 

[jan.didden]

OK, googled and found MathCad Explorer, version 8 is the last available.

BTW a 79 yr old friend once told me, you know you're old if you forget to close your zipper after visiting the bathroom.

You know you're really old when you forget to open it.

You're really, really old if you have a nice nurse to do it for you

Jan Didden

 

[gedlee]

averaging is equivalent to simply using long records to lower the fft noise floor

This is not entirely true. There are ways to better the S/N ratio by synchronous averaging a signal that is synchronous with the time window.

 

[AndrewT]

......... a 79 yr old friend once told me, you know you're old if you forget to close your zipper after visiting the bathroom.

You know you're really old when you forget to open it.

I'm old enough to fail the first test and forgetful enough to not remembering ever failing the second test.
But I'm still young enough to feel the need to put it right.

 

[jcx]

Care to elaborate? - um, I'm referring to Geddes msg, no further input required from jan or andrew on the other subject on my account

also what levels are useful to measure to?

the Juli@ seems pretty good at low levels:
http://www.ixbt.com/multimedia/esi/julia/esi-juli@-2444-4dbu-balanced.shtml

I don't know of any evidence that we can hear anything that requires seconds of averaging to pull out of the (near by in frequency) noise floor - is it valuable to go far beyond audibility limits?


nothing prevents the creation of test signals with integer number of periods in long records to maximize the energy in discrete measurement bins - nor is it required to only look at the magnitude, the fft also includes the phase information in the Re, Im values and this can be used to resolve any component with respect to the any phase you choose relative to your test signal

but interference noise frequency lines can be above the level of the distortion components especially for low level tests I believe we're discussing, and the interfering frequencies will not be synchronous with sample rate or record length, power line noise being particularly hard to avoid

so Blackman windowing can be a good strategy even if it gives up some signal in the smooth tails of the window, the smoothing will decrease the "splatter" relative to the rectangular windowing of interfering frequencies asynchronous with the record length, reducing the interference frequency spectral leakage contribution to the hopefully well separated in frequency test signal and measured harmonics

 

[Conrad Hoffman]

Dang, I've always tried to keep my artifacts buried and hidden by the noise floor. Now you're telling me I have to do better than that? It would be an interesting test to process some music where everything below a given level is zeroed. I wonder how high the zero level has to be, before you hear something wrong? Obviously we already have that at the LSB of whatever digital system is in use, and my assumption is audibility would occur at some higher level. If not, it's back to analog! Pretty sure tests have already been done mixing tones and harmonics down at those levels, though I don't have a reference, and they were found to be inaudible. Thus the statements that anything below xxx% THD is inaudible.

 

[gedlee]

There are some problems with the previous two posts, but they are subtle and I will have to oversimplify a bit.

The idea is to measure the harmonics down to as low an input signal level as possible. This can be difficult as I found out when I first tried it. The noise, interference, etc. got in the way and simple windowing and averaging did not do a very good job of pulling out what I wanted to see.

And yes, the ear can seem to do this even though we have trouble with our instruments. Most of that is because we like the FFT so much (its so easy to use) but the ear doesn't process signals that way. So the FFT usage is getting the way - nothing new there.

But as was pointed out a lot of the problems come from signals that are not sychronous with the time window. My original idea was to lock the input signal to the time window and average in time (the actual time data) prior to taking the FFT. This would have the effect of lowering everything that was not sychronous with the time window even if it was deterministic and not random data. Averaging in the frequency domain will only reduce the random data not the deterministic data.

But, as I said before, the data that I showed appears NOT to do that. It was only averaged in frequency. I don't know why. I need some time to sort this out. I might have tried my idea and it didn't work and then I remembered the idea, but not the fact that it didn't work. I've done that before!!

None of this takes away from the fact that it is the nonlinearity for low signal levels that is important not for high signal levels. I don't know of anywhere where this is reported for amplifiers since, as I said, THD versus level is virtually always THD+noise versus level and this masks the problem.

It may be that we don't need to synchronous average or anything "special". We may just need to look where we haven't looked before.

 

[Bob Cordell]

Crossover distortion is a particularly insideous form of nonlinearity because it happens at all signal levels and there is no comparable mechanism in a loudspeaker to mask it. The question was asked if I have a way of identifying crossover distortion in an amplifier.

Yes, I do.

You see the situation with crossover distortion is that the % distortion increases with falling signal level. This is exactly why it is so audible since this is directly opposite to our hearing.

One could therefor ***** crossover distortion by looking at THD as the signal level goes lower, which is a typical measurement. The problem is that virtually all of these THD versus level measurements are THD + noise. When this is the case, the rise in THD at lower signal levels is actually the noise and NOT the distortion, but it is impossible to tell which is which. SO this test actually masks the real problem. One would have to track the individual harmonics of the waveform, but then the noise floor is still an issue.

Hence the measurement problem is one of noise floor and how to measure distortion products down below this floor.

This is done by averaging. But normal averaging can only lower the noise floor so much - down to the noise power. But if I have a signal and I average this signal sychronously then I can raise the net signal to noise level. This too is common. But if the signal does not exactly fit the time base then I need to window it and the resultant spectral leakage makes this sychronous averaging less effective.

I use a signal that exactly fits into the time base of the A/D taking the data. This means that I don't have to use a window and I can sychronously average a signal to noise ratio that is about 20 dB better than a simpler test could achieve. This means for example that the input signal needs to be something like 976 Hz, not 1000 Hz, which doesn't exactly fit the window.

I actually had to generate the input wav file in FORTRAN using quad precision, special random number generators and rounding techniques, because the test signals needed to have a 120 dB dynamic range - very difficult with 16 bits.

I use a signal that starts out low and goes up in level. I plot out the results as the signal drops into the noise floor. This test shows vast differences in amps that measure identical with standard tests.

It also shows that my Pioneer amp - you know the "really crappy" one that I get crticized for using at RMAF - is an extremely good amplifier. As good as the best that I have tested with this technique.

Hi Earl,

This promises to be an interesting and valuable thread. Distortion measurement has always been one of my obsessions.

You are exactly right about the low-level measurement limitations of THD+N. THD turns out to be a pretty good test, but the noise masking at low levels certainly gives it a bad reputation for finding stuff at low levels.

I also believe that it is important to be able to look below the noise floor. I use a couple of approaches, sometimes all together. I usually employ a very good THD analyzer with very low inherent distortion, then feed the distortion output into a spectrum analyzer. This spectral analysis allows one to look quite far below the noise floor. I happen to use a good old-fashioned analog HP spectrum analyzer (HP 3580A), but a PC-based one like Visual Analyzer with a very good sound card like the Juli@ can also be used.

I also use a device I designed called a Distortion Magnifier (DM). It uses amplitude and phase-matched subtraction of the input and output of the amplifier under test to magnify the distortion by 20 or 40 dB. I used this technique in order to measure the distortion of my MOSFET power amplifier back in 1983. Its described in the paper on my website at www.cordellaudio.com. Built with very low distortion op amps, the DM increases the dynamic range of the THD analyzer (or any other analyzer) by 20 or 40 dB.

Finally, I can use both the DM in front of the THD analyzer and the spectrum analyzer after the THD analyzer to get way down, even at fairly low signal levels. The spectrum analyzer does a very good job of dispensing with the noise (but there is a tradeoff in terms of how long you are willing to wait for the result, especially when very small noise bandwidths are selected). I can see THD below -140 dB with this kind of a setup.

I also use the DM together with the spectrum analyzer (but no THD analyzer) to look at 19+20 kHz CCIF, which is pretty much my favorite test.

I'll look forward to more distortion measurement discussions on this thread.

Cheers,
Bob