[sdiy] Guy researches and talks about how aging components change the sound of a synth

Tue Sep 1 05:41:24 CEST 2026

Though I'd bet $1000 AI wrote that and that the man posting it is doing so as veiled advertising. 

okay NOW I'll stop 

chris 

> On Aug 31, 2026, at 10:39 PM, Chris McDowell <declareupdate at gmail.com> wrote:
> 
> fair enough! 
> 
> Chris 
> 
>> On Aug 31, 2026, at 10:16 PM, cheater cheater <cheater00social at gmail.com> wrote:
>> 
>>> I kind of feel like his thesis feels like AI, too. Like a drum plugin that traces all behavior back to a component drawn on the schematic, that's something claude or chad would think is important
>> 
>> this has been the defining theme in physical component modelling in
>> virtual analog synthesizers since the late 00's. i know people who
>> build some of the best VSTs of that type and this is exactly the line
>> of thinking that was explained to me, sometime around 2010. long
>> before LLMs existed.
>> 
>>> Yeah, old drum machines have lots of variance
>> 
>> he's saying the opposite. they don't have a lot of variance. they
>> differ from the schematic, but they all differ in the same way.
>> 
>>> all of that still reads like I'm criticizing you for sharing, heh, but I'm not! Thank you for sharing. I am criticizing that guy for posting weird AI stuff on reddit
>> 
>> it's easy to fall into the habit of criticizing output that seems like
>> AI to you, but in this case the criticism just doesn't apply.
>> 
>> On Tue, Sep 1, 2026 at 4:52 AM Chris McDowell <declareupdate at gmail.com> wrote:
>>> 
>>> all of that still reads like I'm criticizing you for sharing, heh, but I'm not! Thank you for sharing. I am criticizing that guy for posting weird AI stuff on reddit, which AI probably told him to do to pretend he's not promoting his plugin. "go be active in communities who are interested in your product"
>>> 
>>> Chris
>>> 
>>>> On Aug 31, 2026, at 9:28 PM, Chris McDowell <declareupdate at gmail.com> wrote:
>>>> 
>>>> Well for what it's worth, it does look like AI wrote it ¯\_(ツ)_/¯  the research could totally still be legit. he could have used AI to help with all the research and it's still legit.
>>>> 
>>>> Is he saying anything besides discrepancies in timing and cutoff freqs due to component tolerance? I kind of feel like his thesis feels like AI, too. Like a drum plugin that traces all behavior back to a component drawn on the schematic, that's something claude or chad would think is important. I guess I'm accidentally shitting on your post to the sdiy list by saying I'm not sure this is interesting haha, just my opinion! Yeah, old drum machines have lots of variance. I think -that- is a cool thing about them, totally. it's not mysterious though.
>>>> 
>>>> but what do I know
>>>> 
>>>> Chris McDowell
>>>> 
>>>>> On Aug 31, 2026, at 8:59 PM, cheater cheater via Synth-diy <synth-diy at synth-diy.org> wrote:
>>>>> 
>>>>> Really interesting stuff:
>>>>> 
>>>>> https://www.reddit.com/r/DSP/comments/1w2z13c/every_surviving_tr909_is_40_years_old_what_does/
>>>>> 
>>>>> quoted below.
>>>>> 
>>>>> PS Some people say he might have used AI to write this. Best as I can
>>>>> tell it's not AI written, but what do I know. It seems like the
>>>>> research is legit. Maybe he used it for rephrasing his notes into a
>>>>> post. Or maybe we're seeing a phenomenon where people who use AI a lot
>>>>> just start talking like it. Historical linguists will have a field day
>>>>> on this 100 years from now, in case we survive that long...
>>>>> 
>>>>> -----------
>>>>> 
>>>>> by BeatForge_Dev
>>>>> 
>>>>> Every surviving TR-909 is 40 years old. What does "authentic" even
>>>>> mean    when the schematic and every living unit disagree in the same
>>>>> direction?
>>>>> 
>>>>> I'm modelling the TR-909 the way I modelled the 808 and 606 before it:
>>>>> component by component from the service notes, with one hard rule - no
>>>>> fitted or matching EQ anywhere. Every filter, every time constant,
>>>>> every gain in the model has to trace back to a component that is drawn
>>>>> on the schematic, or it doesn't ship. Validation is against recordings
>>>>> of real units (multiple machines where I can get them), but the
>>>>> recordings are the gate, never the source: when the drawing and a
>>>>> recording disagree, I don't get to turn a knob until they agree. I
>>>>> have to find out why.
>>>>> 
>>>>> That rule is what made the following visible, because without it I
>>>>> would have EQ'd the discrepancies away in an afternoon and learned
>>>>> nothing. As the model got close - close enough that the remaining
>>>>> deviations were a dB here, a few percent of a time constant there -
>>>>> the residuals stopped looking like noise. They grouped. Specifically,
>>>>> they grouped by what the component that owns each parameter is made
>>>>> of.
>>>>> 
>>>>> Time constants owned by tantalum capacitors read consistently long, by
>>>>> 14 to 19 percent, on every unit measured. The 909's tom body decay is
>>>>> set by a tantalum; every reference unit decays 14-19% slower than the
>>>>> drawn RC says it should, and they all miss in the same direction.
>>>>> 
>>>>> A recovery time owned by an aluminium electrolytic reads consistently
>>>>> short. The hand clap has a roughly 228 ms recovery between hits set by
>>>>> an electrolytic; on both units I could measure, the effective
>>>>> capacitance comes out around 0.83 times the drawn value. Electrolytics
>>>>> dry out and lose capacitance - this is the single best-documented
>>>>> aging mechanism in the business, and it points exactly this way.
>>>>> 
>>>>> A clock frequency owned by a carbon-composition resistor reads
>>>>> consistently high, by about 16 percent - on four different machines.
>>>>> Carbon comps drift upward with age, heat and humidity; four units
>>>>> agreeing on the direction and roughly the magnitude is not a lottery.
>>>>> 
>>>>> And the control that makes this believable: not everything drifts. The
>>>>> kick's poles sit dead on the drawn values across units. Film caps and
>>>>> metal film resistors holding their values while electrolytics sag,
>>>>> tantalums stretch and carbon comps climb is exactly what a 40-year-old
>>>>> PCB should look like. If every parameter had missed, I'd have
>>>>> suspected my own readings of the schematic. The residuals sorting
>>>>> cleanly by dielectric and resistor construction is what flipped my
>>>>> interpretation: these aren't model errors. The model was measuring the
>>>>> age of the reference units.
>>>>> 
>>>>> Which lands you somewhere genuinely strange for validation. The
>>>>> schematic describes the machine Roland built in 1984. Every recording
>>>>> anyone can make today describes a machine that has been drifting away
>>>>> from that schematic for four decades - and drifting coherently,
>>>>> because the same chemistry is running in every unit. So a model that
>>>>> is exactly right by the drawing will sit at a small, systematic,
>>>>> same-signed offset from every reference you can buy or record. If you
>>>>> validate purely by fitting to samples - which is the industry default
>>>>> - the fit will happily absorb one particular elderly unit's drift,
>>>>> your metrics will reward you for it, and you will ship "authentic
>>>>> 1984" with today's electrolyte chemistry baked in.
>>>>> 
>>>>> Edit:
>>>>> 
>>>>> After listening more and more, i prefer the old values, Toms and Kick
>>>>> have more sustain, it just sounds more alive and less stiff.
>>>>> 
>>>>> Question for the room: has anyone treated component aging as a
>>>>> first-class model parameter rather than a nuisance? Tape emulations
>>>>> have had "worn" controls for years, but those are usually designed
>>>>> sounds, not chemistry-driven parameter drift. And if anyone has good
>>>>> literature on solid tantalum capacitance drift over multi-decade
>>>>> timescales, I'd genuinely like to read it - the electrolytic and
>>>>> carbon-comp stories are textbook, but the tantalum shelf is thinner,
>>>>> and right now my 14-19% is a measurement in search of its chemistry
>>>>> paper.
>>>>> 
>>>>> Disclosure: I make a commercial drum plugin, this work is in it.
>>>>> ________________________________________________________
>>>>> This is the Synth-diy mailing list
>>>>> Submit email to: Synth-diy at synth-diy.org
>>>>> View archive at: https://synth-diy.org/pipermail/synth-diy/
>>>>> Check your settings at: https://synth-diy.org/mailman/listinfo/synth-diy
>>>>> Selling or trading? Use marketplace at synth-diy.org
>>>> 
>>> 
> 