Amplifier distortion testing using a modestly priced audio interface

[AnalogSteph]

Not too surprising. FlexASIO is a wrapper so can fundamentally take most anything, not to mention it's a WASAPI wrapper and WASAPI internally operates on float32 before converting that to whatever the output device might support. Hence, FlexASIO will take up to float32 by default (which you can apparently change if desired).

It is common for drivers to accept int32 even if the hardware is 24 bit only like yours, just because 24-bit samples are majorly annoying to deal with in a world where everything is powers of 2 in length. The last 8 bits are merely filled up with zeroes (zero-padding). Hence why REW has a "Treat 32-bit data as 24 bit" checkbox.
I still vaguely remember using 32-bit MME recording with one particular card back in the day, and more recently in RMAA to circumvent some (possibly CMedia / Asus Xonar) sound driver bugs.

BTW, the L at the end stands for little-endian, which is the prevailing byte order in x86 land.

 

[DavidR]

I started this thread for use with the Scarlett 2i2 that I have, but much should apply to others. I say that because I found similar issues with the 1010LT PCI card. I'm also still testing the 1010LT alone and the 2i2/1010LT balanced combo, so I guess this is more of a generic testing discussion. Many old PCI cards are actually still very good as are some (really expensive) new cards. I'll be posting more results in that vein.

 

[DavidR]

BTW, the L at the end stands for little-endian, which is the prevailing byte order in x86 land.

Ah, I wondered about that. Had some experience with that many years ago when trying to "see" data values in a data dump. Painful.

 

[AnalogSteph]

Many old PCI cards are actually still very good as are some (really expensive) new cards.

That's pretty much been the theme with better-quality PC audio equipment in the last 20 years or so: You don't replace it because it's no longer fit for purpose, you replace it because it no longer works with your new system or is too much of a pain to use or defective. Just look at how long the old E-MU 0404 USB from 2007 remained a lab staple, easily 15 years.

 

[DavidR]

Similar to earlier tests of the 2i2/ADCiso loop, I've done more testing of a 2i2/1010LT balanced loop with somewhat surprising results.

One thing to point out is that for the Scarlet 2i2 and the M-Audio 1010LT noise dominates.

The first measurement I ran, as always now, was the Stepped-THD vs Level with a 0.1dB step. These are very much a jumble across the bandwidth, especially if higher HD components are graphed. It's impossible to find the sweet spot with this. It stops at HD7 because higher HDs are low enough in the desired range. They do, of course, get included in the THD calculation.

So the next thing I always do now was to expand the graph to open up the likely usable range. This provides a view to all HD components clearly.

After much trial-and-error testing I settled on REW Gen output at -8.67dBFS -11.4dBFS This set the 1010LT balanced input at -11.40dBFS -8.67dBFS. Note that just 0.1dBFS change can be significant in results. Some HD components may go a lot higher while at the same time other components go lower. It's a balancing act finding the "sweet spot". If one is only interested in, for example, only HD2 and HD3 then the sweet spot may very well be very different than for other tests. Noise changes for almost every point, thus so does SINAD. This is one reason why I see little reason to focus on SINAD.

Edit: Had In/Out levels flipped.

The numbers in this graph are for the Coherent measurement, but the Standard results were very close, Coherent slightly better of course.

For comparison here again is the 1010LT Unbalanced Loopback.

Two things to note. First, they're not that much different, more for individual HD components and noise. Higher HD results were actually better for the unbalanced loopback.

Second, the "sweet spot" is very different between the two. I can only think of two possible reasons. One is that the 2i2 output is significantly different than that of the 1010LT. The other may be related to the 1010LT unbalanced loopback allows REW to use the rectangular window due to a single clock. I don't know how much that affects results. Still, the 1010LT unbalanced results are excellent in my opinion.

To emphasize the point about different sweet spots for 2i2-1010LT balanced vs 1010LT Unbalanced this graph shows the 2i2-1010LT Balanced measurements with the input to the 1010LT set to the same voltage (dBFS) that I selected as the 1010LT unbalanced sweet spot. The results are significantly different that emphasizes the difference between output of the 2i2 vs 1010LT. Of course this is a balanced-to-unbalanced input comparison, so take that for what it's worth.

The 1010LT appears to be more than sufficient for testing unbalanced amplifiers, noise aside.

 

[DavidR]

This may be a moot point since there is no unbalanced output in the 2i2. I connected the 2i2 to the 1010LT unbalanced input to see how it compares to the 1010LT unbalanced loopback. The 2i2 is totally unusable for this. The one word that comes to mind is...awful. There's no usable "sweet spot" at all. What is puzzling to me is that the issue is entirely HD2. Higher even-order and all odd-order have an area that would be usable as a sweet spot, but the HD2 is never better than -85dBFS of the 1010LT input and that's in a very narrow bandwidth. Noise is much worse as well, but I suspect that's due to the same problem as HD2. It's not worth posting any measurements of this.

This emphasizes how much better and useful the PCI 1010LT will be for unbalanced use.

Also, it means that if one will want an audio interface for balanced and unbalanced testing, the 2i2 is nowhere near adequate.

 

[DavidR]

I have some results of testing one channel of a Kenwood KX-M1 6-channel amp, one of three I have, that I adjusted for bias and offset. Two amps had all six channels with offset 0 volts, no adjustments needed. The third one had all channels bypassed, hard-wired, and needed serious bias adjustments, but was adjustable to spec. The amp tested still needs output relays replaced, but the bad ones are bypassed for now. This test is for a bypassed-relay channel using the 1010LT unbalanced I/O and REW. The results are excellent, though just a first set of tests. Testing is limited to 1W and 5W because I don't have any high power resistors (yet). The reference documentation by Kenwood for distortion is 65W into 8 ohms, stereo-mode, so that's what I plan to do at some point. It doesn't say both channels driven. I plan to only test with one channel driver since this is a six-channel amp and I can't see trying to testing with multiple channels powered. It should still be an interesting test of an amp more than 30 years old. The service manual is dated 1993.

I found that using the S-THD from the 1010LT unbalance loopback didn't provide the best "sweet spot" once the amp was being tested. I suspect that this is partly due to the difference in the sound card output because it has to be set to a specific level to produce the required input voltage to the amp for 1W into 8ohms for this test. I then measured the amp output voltage into the 8 ohm load and adjusted REW output until the amp measured 2.83V. This provided the output dBFS value in REW for 1W. Next I ran a Stepped-THD vs Level for the amp. This was safe to do by setting the REW S-THD End value to that 1W dBFS value. This provided a set of distortion curves so that I could try to find the optimum input dBFS (voltage) into the 1010LT input, as I have done for audio interface testing. That result is shown below.

It looks pretty good down to about -4.5dBFS, but as the saying goes, the devil is in the details.

This is the expanded graph to be able to see needed detail.

Using this I settled on -4.2dBFS, although I tested a number of other points.

I should also add that the feedback probe goes back to a linear 10-Turn pot that allows me to precisely set the feedback level. I first have to determine the values (empirically) to set the level so that allows the feedback to go a little above the desired 1010LT input a bit above -4.2dBFS. With the 10-turn pot I can then set the dBFS level precisely to two decimal places of dBFS. I'll need multiple probes for testing different power levels. 1W required no probe because 2.83V is below the -4.2dBFS optimum, so that low of amp power level required maximum feedback given that the 1010TL input 0dBFS is around 4V.

 

[DavidR]

Next are the RTA results for -4.2dBFS, 1W and 5W.

I also ran REW with the Coherent setting for 5W. There is some improvement in the HD levels. What surprised me was the nearly total elimination of all 60Hz related signal and all of its harmonics.

Just as for the Scarlett 2i2 and the 1010LT card, noise dominates SINAD, but that doesn't concern me. I've used these for years and never heard any influence of noise in my speaker systems. They aren't very sensitive systems, though. This does reinforce my thoughts that I don't need to replace these 30+ year old amps with something newer (and seriously expensive) since I've been able to perform adequate maintenance. No cap changes planned at this point, either.

By-the-way, the Kenwood spec for 65W into 8 ohms is THD 0.0015% (-96.5dB),

 

[DavidR]

Here is the Multi-Tone S-THD test. The REW output setting was the same as that for the single-tone test. However, the ten-turn pot was used to lower the input to the 1010LT. The single-tone sweet spot doesn't work for the multi-tone, so this required some trial-and-error changes to get this result.

 

[DavidR]

I purchased a pair of low inductance 8-ohm power resistors for amp testing. These came from Parts-Express. After a few RTA measurements I ran an S-THD for one 8-ohm and again for the sand cast resistors immediately afterwards that I had been using. The amp had been powered all day, though at idle most of the time. There shouldn't be any difference due to lack of any "warm-up". Both measured 8.0 ohms in my DVM, but it only shows single digit decimal point.

I was immediately struck by the difference in Noise Floor between the two. I had not expected that and not sure what to think about that. The difference is about 3.5dBr for the low inductance resistor. This graph is for the ADCiso input. I may do some re-testing to see if I made an error somehow, but. I had edited the measurements incorrectly. The graph below is correct. The only change I made was swapping resistors between measurements. I'm still not sure what to think about the difference. This would imply that there's a serious difference between these resistors. I still have to examine and compare the individual HD components.

Edit: I re-measured. The results were identical.

I'll be posting some RTA comparisons as well. The one above is and new ones will be for the 1010LT out to E1DA ADCiso in measurement. That surprised me as well. The Scarlett 2i2 was unusable for unbalanced amp testing. Despite being unbalanced, the results with the ADCiso are a bit superior to the 1010LT I/O measurements, although not by a lot. I'll have more to say about that later with some RTA comparisons.

Edit:
I've done some comparison between using the Monitor1 vs 10-Turn Pot between the amp and ADCiso. There is some benefit to the Monitor1, possibly due to the Monitor1-to-ADCiso being a balanced cable.

There also doesn't appear to be any real advantage to using a low-inductance resistor in place of the sand cast ones. I can only guess that it's because this amp doesn't have the much better specs of recent amps.

 

[DavidR]

Kenwood KM-X1 6-Channel Amp, Channel 1, RTA of Sand Cast vs Low Inductance Resistor, 48kHz 256kFFT Coherent. Delta 1010LT output to Behringer Monitor1 to E1DA ADCiso Grade A input. Definite improvement in most HD components. This is with the Monitor1 at Max Dial setting, nearly identical to direct to ADCiso. Red is the Sand Cast resistors, black is the single 8-ohm low inductance resistor.

 

[phofman]

I also ran REW with the Coherent setting for 5W. There is some improvement in the HD levels. What surprised me was the nearly total elimination of all 60Hz related signal and all of its harmonics.

Since that 60Hz signal from the power net is not phase-coherent to the fundamental signal from the generator, the vector averaging (AKA coherent in REW) suppresses the bin. Vector/coherent averaging is designed for analyzing harmonic distortions and other components coherent (phase-aligned) to the fundamental.

 

[DavidR]

Since that 60Hz signal from the power net is not phase-coherent to the fundamental signal from the generator, the vector averaging (AKA coherent in REW) suppresses the bin. Vector/coherent averaging is designed for analyzing harmonic distortions and other components coherent (phase-aligned) to the fundamental.

That makes sense. That would also explain the reduction in the Noise Floor as well.

 

[DavidR]

I have a question. What is the reason for the more accurate measurements when output and input are on the same device, therefore the same clock? How does the clock timing affect it? I know that within REW it allows the use of the rectangular window, but I can't see why the clock is so important when the clock rates are so high.

 

[phofman]

IME coherent averaging also reduces small coherent harmonics/distortions where the noise in the FFT bin makes the phase of the distortion slightly deviate against the fundamental. Typically measurements of low distortions look better with the coherent averaging than with the standard amplitude averaging (amplitude may not change but the phase does due to the added noise does).

 

[DavidR]

IME coherent averaging also reduces small coherent harmonics/distortions where the noise in the FFT bin makes the phase of the distortion slightly deviate against the fundamental. Typically measurements of low distortions look better with the coherent averaging than with the standard amplitude averaging (amplitude may not change but the phase does due to the added noise does).

Interesting. I've noted that generally the HD components do appear to be much lower for coherent averaging (not without exception). But if the phase is the culprit, that raises the question, which measurement is more accurate? The noise alters the phase in an HD bin, therefore the reported magnitude (makes sense), so the non-coherent measured result would have errors due to the noise phase. If so, why not always use coherent measurements?

Another question. Does averaging reduce the impact of the phase issue since it's random against the fundamental and harmonics?

I've turned on the reporting of phase for HD components before, but I can't make sense out of it for interpretation or clarification of accuracy.

 

[phofman]

I've noted that generally the HD components do appear to be much lower for coherent averaging (not without exception). But if the phase is the culprit, that raises the question, which measurement is more accurate? The noise alters the phase in an HD bin, therefore the reported magnitude (makes sense), so the non-coherent measured result would have errors due to the noise phase. If so, why not always use coherent measurements?

The "legacy" averaging uses only amplitude if the bin and does scalar averaging. Whereas the coherent averaging does vector averaging. IIUC the amplitude of vector average is always lower or equal to the average of amplitudes only (bcs any triangle side is always shorter than sum of its remaining sides).

IMO the choice depends on the needs. IMO the vector averaging is good when you want to fish for the harmonics hidden in the noise as the uncorellated noise will be elimininated more by the vector averaging. But for considering the device parameters I would use the amplitude average as simply rotating a distortion of equal amplitude by some angle does not make it any smaller (which the vector average yields).

 

[DavidR]

I made a number of tests of the Kenwood KM-X1 using different input hardware to find the best one for testing an unbalanced amp. All used the 1010LT unbalanced output because the 1010LT I/O loopback results were so good. Good enough for my needs, but since I have the Scarlett 2i2 Gen 4 and theE1DA ADCiso I was curious to see how well those did as the amp feedback input. The results have been surprising. Again. However, I never got the 2i2 to work with the 1010LT output. FlexASIO kept failing. The ADCiso was fine, however.

These were all run using the 8-ohm high-power low-inductance resistor pictured in Post #70. The feedback was fed through either the Behringer Monitor1 or the 10-Turn pot when input to the Delta 1010LT. The ADCiso was connected directly to the amp due to the drive setting for 5W amp output. This kept the voltage a bit under the limit of the ADCiso for the 6.7V setting.

I also run the S-THD through the amp with each feedback device, setting the stop point for 5W into 8-ohms (6.32V). This let me find the best input level, the same as I did for the audio device loopback tests. The RTA was then run based on this. Let me emphasize again that using a probe feedback with a somewhat random divider ratio will very likely not result in the optimum audio interface point (sweet spot). Any HD value can easily be +/- 20dB different between any two selected dBFS (voltage) points on a device.

Here are the results of each input device, Legacy vs Coherent. The numbers on each graph are for the Coherent average. They are in the order that I consider worst to best, although they are all certainly acceptable for my needs. I didn't choose the order based on SINAD (THD + Noise). That is nearly identical on all due to the noise of the amp which, by today's standards, is not very good. But I use one of these for a three-way dipole system and another for design speaker testing and don't notice the noise. This is using the Ultimate Equalizer from Bodzio software with all drivers directly connected to the amp, not even with a safety cap for the tweeter. A tweeter has never blown.

I base the order primarily on HD2 and HD3, though emphasizing them is probably debatable. HHD in REW includes all up to HD50, so it's obvious in testing that at times the higher HD components are influencing HHD a fair amount at times, but my thinking is that those aren't very important in comparison to the low ones.

First up is the 1010LT Balanced input. The balanced input, of course, requires an unbalanced-to-balanced cable for the probe connection. These I made. This had to use a probe with a balanced voltage divider due to the 1010LT input 0dBFS of 4.0V. The probe was input to the 10-turn to set the level for what was closest to a sweet spot.

Next is the ADCiso Grade A, direct connection. This may be limited somewhat by having to use FlexASIO and the Blackman-Harris 7 window whereas the 1010LT-only loop allowed use of the Rectangular window. No way to know the absolute impact on the HD numbers.

Last, and best (surprisingly), is the 1010LT Unbalanced input. This also had to use a probe with a voltage divider.

This shows the results of the Legacy measurement. It's curious that even-order HDs tended to look better while odd-order tended to look worse. I generally ran the RTA for the Legacy first, immediately followed by the Coherent, so no time delay.

 

[DavidR]

There are several takeaways from all of this as I see it with regard to using a "modestly priced audio interface". Some of this may extend to higher (read more expensive) ones as well.

One, noise is a distinctly limiting factor in this amp, at least in the numbers game, specifically SINAD. Having used it for many years, I can't say that it the inherent noise had an audible impact.

Two (speculating on this), when measuring an unbalanced amplifier, an unbalanced audio interface may have an advantage over a balanced one due to the common ground of output and input.

Three, the common clock of output and input allows REW to be used with the Rectangular window. The absolute impact of this is not clear nor obvious.

Four, a potentiometer can be used to control the feedback signal for the "sweet spot" in the audio interface. In my opinion and (limited) experience, it may even is beneficial and may even necessary to adjust the feedback for a pre-determined sweet spot. The caveat may be that this is partly due to the quality of the signal source, i.e. the audio interface output. Noise and distortion in the source will almost certainly impact the distortion measurements. Either a commercial one, such as the Behringer Monitor1 for balanced use or a linear 10-turn pot for unbalanced use is recommended.

Five, at higher power levels a feedback probe with voltage divider(s) will be necessary, but that should still be used with a potentiometer that is used to adjust the feedback for the input sweet spot.

Six, a low inductance resistive load evidently provides improved results, more so than I expected.

Seven, I found that the "sweet spot" for an audio interface changes when an amplifier is inserted in the loop to be measured. I had assumed it would be the sweet spot determined in the audio interface loopback. Not so. What I had not considered is that the sweet spot is a combination of output and input. When an amp is inserted, its gain may (likely will) make that sweet spot change because the driving signal, for example a 5W amp output required for a test, requires a fixed audio interface output signal to the amp. This output may not be anywhere close to the interface loopback sweet spot output level previously determined. This is the reason for running a limited S-THD of the audio interface/amp combination. REW makes this a simple, relatively safe operation. Caveat: Take extra care. Damage to the interface could occur if the feedback signal is not kept within the acceptable range of the interface.

Eight, probably obvious, comparing measurements using different test rigs is problematic.

 

[Blumlein 88]

If you really do more than occasional amplifier testing, probably the best option for those without an AP rig is the current Quant Asylum device. One can work around quite a bit using an audio interface by various means. The Quant Asylum makes it all pretty easy and it is capable of getting some good results. It still may not be good enough to probe the full limits of the very state of the art, but it gets close enough that once something is running up against its limits it is good enough not to worry about at all.

QA403 Audio Analyzer

2-Apr-2025: The QA403 is back in stock Affordable and Powerful Audio Analysis The QA403 is our 4th generation audio analyzer. Since 2012 we've been providing audio test and measurement to companies around the world. And with each iteration, the capability and programmability of the...

quantasylum.com

QuantAsylum. Innovative tools for audio engineers.

Since 2010, QuantAsylum has been developing unique products for engineers working in unique spaces.

quantasylum.com