Memory Distortion

I’ve been reading Arto Kolinummi’s book about audio amplifiers. He devotes a lot of pages to the subject of memory type distortions. I’ll quote him from his book:

“This type of distortion can be characterized as memory distortion because it is based on stored energy which is released with some time constant causing an error after the offending signal has passed through the system. Energy can be stored in various ways such as heat, charge, or mechanical energy. Memory distortion error can be totally invisible in distortion tests with steady state input signals.”

I’ve been trying to figure out how to measure this.

One approach would be input a pulse of some kind and then measure the resulting harmonic spectra over time. That’s how a lot of loudspeaker measurement programs work. They call the output “cumulative spectral delay”.

That would be a nice solution, but I’m not sure how to do that with a QA401.

Another approach might be to capture the peaks of a spectral plot, where a lot of cycles are captured at their peak, instead of their average. Kinda like a Max Hold or Peak Hold on a spectrum analyzer. The opposite of averaging. (I think it’s the opposite…) This might also be valuable in measuring DACs, since some of the processing within a DAC is hardly synchronous with the applied tones, so various products might appear on a random basis that average out over time.

This probably means that a different software detector or faux window process is needed to capture and hold peaks.

The function is described here: http://www.w7ay.net/site/Applications/Amici/index.html

Are these functions that could easily be rolled into the next great rewrite of the software? Please? Or, will the new platform be flexible enough for somebody to add these on their own?

Hi @BkDad, It seems the key to this analysis is the time constants involved. Speakers responses have time constants on the order of milliseconds, so there’s enough there to resolve with FFTs at typical audio sample rates. But as you shift to other components, the time constants are likely much, much shorter, meaning you’d need scope-like sample rates.

You can hit just about anything with an impulse response to see what the response and spectrum of that response might look like. If you wanted to do some experiments with the QA401 hardware, you could use the ASIO driver and then run ARTA’s spectral decay software.

I do think there’s a need for some software to help characterize and understand temporal aspects of systems. Mostly, I’d like to be able to measure with great accuracy the frequency-dependent delay processing through various effects. But not sure how that might overlap with the above.

I’ll give ARTA a try. Great idea! Thanks.

I still think that a peak hold detector would be useful. I think it’s “merely” just taking the peak value from each FFT bin and displaying that. (I say merely, because I’m not the one doing it!)

Issue added to git

Wow! Thanks!

It appears that ARTA will do the cumulative spectral delay measurement. In fact, it does it in so many ways that some experimentation will be required. But, that’s part of the game. Great suggestion!

One more thing, unrelated.

I see that ESS now has some pretty good looking ADCs to go along with their DACs. Although it would be a PITA to change, at least they didn’t have a fire at their chip factory. The ADCs look to offer some good advantages over the AKM parts. On paper, anyway.

Please report back what you find. My hunch is that the “memory distortion” you will find on passives (likes Rs and Cs) will be so fleeting that FFT at audio sample rates won’t find them. You’ll need many MSPS and 16-18 bits of resolution to see them. If that is true, then the inertial aspects of the decay associated with speakers will dominate by a factor of 100 or 1000. But that’s just a guess.

I will. It’ll take a while. Slow work takes time…

Kolinummi points out that large caps have a problem, but that thermal effects are even larger. He claims that in some ways, the waveform distortion can be greater due to thermal effects than for non-linear distortion. He has some measurements, but more are simulated results.

Yes, agree, power dissipation in a resistor divider combined with a lousy tempco can show bad distortions at low frequencies (and the resistors warms and cools with each cycle) that tend to go away at high frequencies as the temp converges to a steady state.

That fellow from your local “competition” (Bruce Hofer) has published a paper that describes why Vishay foil resistors have higher distortion that’s measureable at lower frequencies. All due to thermal effects.

Kolinummi shows that it’s worse for semiconductors. There’s a large difference depending on the device package. I guess that can be a function of PCB trace layout then, too. The IC designers I think figured out how to deal with this a couple decades ago with various layout tricks.

The resistors in question, Vishay Fil resistors will be sensitive a very low audio frequencies like 20 Hz or below because they are made from two elements with complimentary thermal characteristics. At low audio they don’t track as well but the distortions are still incredibly low. The thermal distortions do exist in semi’s but feedback and careful thermal management can really reduce them. The time constants for the thermals can be short but most are pretty long. An easy way to look at them is settling time measurements. Those are difficult to do since you need a clean transient source and usually a clipping amp with really fast settling (tek 7A13 has been the usual device) so you can see the .01% wiggles in the tail.

I’d considered building a Jim Williams type settling measurement system. But, since fooling around with audio is something I do for fun and don’t get paid for, I’d rather spend the same time laying out a PCB for a preamp, as an example. Maybe I’ll still need to go that route if ARTA doesn’t provide any insight.

Anyway, Dr. Kolinummi makes a pretty good case about thermal distortion and offers some remedies for same. You could also consider a similar issue with charge removal in multiple stage emitter followers when the driver turns off. It’d be nice to measure it.

All of this might not matter, but then again, my experience has been that amplifiers often sound different from each other and that doesn’t always track with steady state distortion measurements. That would be nice to get some insight into. The good part of this being a hobby is that I can go down whatever rabbit hole I choose to.

1 Like

The good part of this being a hobby is that I can go down whatever rabbit hole I choose to.

Indeed it is. And those are the best itches to scratch too, because you become can become an expert in the process.

1 Like

Ha! I don’t think the word expert will ever be used to describe me with regard to anything. I’ll settle for “modestly less ignorant.”

The C of transistors is a cause of distortion and the C value varies with the voltage level. esp b-c capacitance or g- d capacitance. Certain circuit topologies can cancel most of the C. Thereby, not depending only on feedback to reduce distortion.

RNMarsh

1 Like