Thanks for this explanation, Unaspected. Better than mine! And I’ve coded a few deep neural networks myself...Unaspected wrote: ↑Thu Oct 22, 2020 9:42 pmThe task would be beyond daunting if you were to approach it as a human using this method - Also noting that a human has the advantage of a complex understanding pitch before he approaches the synth. This isn't tuning an additive bank of sine waves via FFT; this is tuning each partial/oscillator by trial and error untill best guesses are arrived at - whilst also tuning every other aspect of the synth and calculating how to play it so it best matches the input data.fmr wrote: ↑Thu Oct 22, 2020 3:11 pmThat's where your line of thinking fails. Synthesizing the sound isn't "that" difficult. Actually it was done in the sixties by Jean-Claude Risset using acoustic analysis and additive synthesis via the Music language created by Max Mathews. It has nothing to do with human limitations, since we can program each element at a time, and then trigger most of them using relatively simple controls.perpetual3 wrote: ↑Thu Oct 22, 2020 2:52 pm .../... if the sound is complex enough, eventually it becomes really difficult, for nearly everyone, to synthesize a physical sound using elementary modules or DSP elements because humans have limitations (two eyes, two arms, ten fingers, difficulty paying attention to more than one thing at once, etc).
This is why we have sustained sounds which seem to be made up of scraps of audio. The machine hasn't finished learning but that is the best match at that point.
Perpetual3 does seem to understand the process well.
If we take Gaussian distribution: We could start with a cluster of sound - seemingly random - but as time progresses, the sound converges on a single pitch and floats around that. The difference is that whilst an orchestra generally understand their instruments and have a reasonable grasp as to the starting and end point, the machine only knows a reality of what previously worked and what didn't.
As a slightly meta example: Consider how the Fourier Transform uses correlation: Except, imagine the machine having no algorithm. So it prints each sample value by sample value, making comparisons to the original audio, keeping the best matches and scrapping the worst. It would be a much slower process in human terms. However, at some point, we might find that the machine learns how to generate and match sine waves. Within that data there will be some understanding. Just as with Google's audio engine, the machine will work out tones and articulations that sound realistic.
Consider Google's offering of proof of concept - and it sounds damn good to me. Very impressive work with plenty of possible audio applications for both generating and processing.
