I’m reasonably sure that analog and electronic keyboards could be “unmixed” down to individual parts, but a whole lot of model training would be needed for this given the essentially unlimited variation in timbres that exists, along with some kind of ability to discriminate electronic parts that don’t correspond to well defined categories like “acoustic guitar” and “piano.” But. Is this on the horizon somewhere?
1 Like
touche!
Depends on how one defines “horizon”. A horizon that’s a year out? A decade? 
Your hunch regarding what’s involved is correct.
I won’t claim to speak for SL or all the hundreds of 3rd party guys who do underlying training/model code & tweaks for demix tech but, I’m sure you understand the following situation…
…A mixed mono audio track consisting of a grand piano and 1975 Polymoog that’s playing various chords/single notes.
…A mixed mono audio track consisting of a rhodes, mellotron, Prophet 5, harpsichord, and Jupiter 8.
…A mixed mono audio track consisting of a B3, Arp 2600, upright piano.
…A mixed mono audio track of Stevie Wonder rhodes and clavinet only.
And so on.
As you can understand, a Spectralayers “button” or buttons labeled “acoustic piano” and/or “synths” etc etc…when invoked…will…fail to do the expected.
Why?
Because how in the world does one create a few lines of code (circa 2025 anyway) to have learned every possible tonal combination of just the tiny subsets listed above?
This capability you ask for will be…imo…fairly manageable in a decade or so when off-the-shelf programs have more on-the-fly-train-this-specific-2soundSource-combination-and-then-demix-them-while-I-wait capability…rather than where current state of the art is after a mere five years of the tech being available.
Now…if you hire a 2025 crew to specifically pull YOUR specific Matrix12+wurlitzer piano off a mixed mono track…ie…those guys set up training parameters for that specific combination…you just might get a good result after investing a few thousand dollars…although…“that” code will be completely unusable for your track that happens to be your Matrix12+Farfisa.
In the meantime, a couple of presets that…circa 2025…essentially fail miserably when invoked on soooo many varying combinations of timbres/sonics? No thanks (if me). People have unreasonable expectations of what the tech can do and I for one…wouldn’t want to tick off 60% of the crowd for which the code-chosen-preset is useless most of the time.
Reminds me of the funny line in an “Anchorman” movie…“tests have shown that in 20% of cases, it works 100%”. hahahaha!
Oh well. Someday
1 Like
Well, some things could probably be trained without extraordinary effort, for example:
Clavinet
E Piano
Synth strings
Tonewheel organ (clean, distortion, etc)
Pipe organ
Sawtooth/sync lead
Acid bass
there is just too much timbre crossover
even if the algorithms were trained on the solo instruments; unmixers might experience some untangling; but I don’t see it happening…only the the really obvious solo parts…
consider voice recognition…then how about unmixing multiple human voices? with (much) more distinctive timbre differences, we still experience plenty of sound sources information the algos cannot differentiate
then add in the keyboards having essentially the same voice characteristics,
and…
the unmixing of near perfect solo stems seems very unlikely
I will be very interested to see technology prove me completely wrong