TLDR: Skip to proposed testing methodology and/or knowledge gap(5).
The question "what makes one cap (or resistor) sound better than another" is one of the questions I wanted to solve when I bought 2 S.E.X. amps. And now that I am rebuilding those amps I thought I would revisit that question with the addition of nearly 20 years of experience.
The original experiments I did, much like others, were resounding failures. There was simply no discernable data to be had taking individual, uncorrelated measurements with a multimeter or cheap old scope. I definitively determined that motor run/start caps sound like trash. But I didn't get any closer to why.
Hypothetical reason related to every weird snake oil advertising claim can be found. But nothing to objectively back anything up. No repeatable data that can be used predictively at least. Just anecdotal or group agreement that x material or brand sounds x way. Which doesn't sit well in my brain.
So being a systems integration and robotics specialist, and definitely not an analog or audio circuit designer I sat down and had a good think about how what I know, and am good at, could solve this question. And I think I may have a testing methodology that could provide useful data here.
In robotics one of the big problems is the tendency for free moving effectors to be herky jerky (yes that's the technical term). And similar group agreement, without data, predominated how to address this with roughly the same result. Not much.
Until a systems integration approach was taken to data collection.
Applied to analog circuits my hypothesis is this: That by using multiple scopes, data logging multiple attributes of the primary circuit points and then using rudimentary machine learning analysis to analyze that data, it should be possible to determine the effect that changing components has on the system. And most importantly, that the testing also needs to include fully correlated spectral and electrical analysis of a driven speaker.
This way the absolute values become moot. Only the difference between subsequent values are relevant. And as long as the test equipment has sufficient resolution to measure those differences and those differences are repeatable...a model can be built to predictive accuracy.
And given the advances in affordable measuring devices, hopefully it can be pulled off using affordable, prosumer gear such as Digilent Analog Discoveries, Umics, REW, Dayton or similar digital speaker testing equipment and custom trained open source AI.
Proposed (very loose) methodology:
Use multiple AD 2 digital scopes to simultaneously data log time correlated circuit conditions for each component under test and each major circuit block.
Construct a front and rear sealed anechoic speaker box with calibrated mics at fixed locations. Spectral analysis via REW. Electrical analysis via DATS.
Data crunching via a custom LLEMA model trained specifically to deal with the resulting data streams, formats and circuit topology.
Known knowledge gaps:
(1)The measurement points most likely to reveal useful data.
A brute force approach would require a pile of 40ish AD2s, laptops and a decent server to coordinate them. And that's more than I want to spend on a proof of concept. I'd prefer to be around the 5ish AD2 mark for both cost and setup complexity. At least at the proof of methodology stage.
(2) Has anyone attempted anything similar? I wasn't able to find any similar testing conducted in the peer reviewed audio world.
(3) Are tube circuits temperature dependent enough to require not only time correlation but temperature correlation as well?
(4) Are tube stages stable enough to result in repeatable tests?
(5) If the equipment is sufficiently precise to measure differences, can those differences be correlated to something currently understood, measured and reported by the component manufacturer? And if not, is it something that can be measured about the component itself outside the circuit?
(5)Do I simply have too much time on my hands that would be better invested in learning how to carve bears out of logs with a chainsaw?
