I have a firm grasp on genetic programming, PSO, trees and search strategies like Minimax, etc. I haven't yet got to the point of learning about neural networks beyond the absolute basics.

A project I was humoring involved audio signal processing. I want to know viable AI strategies to implement what I was thinking.

Consider two musicians with two different instruments. The first is playing a violin and the second a guitar. Even if both played the same note, they sound different because of the shape of the instrument, perhaps string gauge, even technique. I wonder of possible ways to make the guitar sound like the violin or vice versa.

Given an audio recording of a violin playing a scale, then a live performance of a guitar playing the same scale, what AI methods could shape the output sound of the guitar to be close to the sound of the violin?

I think it would boil down a lot to equalization. Different frequencies can be boosted or diminished, but how do you find the perfect equalization settings? Hence the question.

What AI strategies would be well suited to this, specifically something that can be done on the spot? In my experience, tree searches take time and might be inappropriate. Genetic algorithms may be better. What about neural networks?

I already have an obvious heuristic: how closely the sound waves match. I just wonder about speed.

  • 2
    $\begingroup$ Not an answer but a pointer: Look at the "deep fake" applications. There are several which imitate specific people's voices, I'd research how they are doing it, the same algorithms may be applicable to your question. $\endgroup$ – Hans-Martin Mosner Nov 8 at 5:54
  • $\begingroup$ "I already have an obvious heuristic: how closely the sound waves match." You don't have this heuristic in production, only for training with deliberately-produced samples. By itself, this points to some kind of machine learning, in order to create any heuristics - or simply the end result - in the absence of any training target. That doesn't rule out searches, such as GA, as a mechanism to drive the ML. $\endgroup$ – Neil Slater Nov 8 at 8:46
  • $\begingroup$ Question for clarification: Can you describe your target system? There is a big difference between switching instrument sounds in a controlled environment, e.g. a clean monophonic recording of a single instrument, compared to in "the wild" e.g. during a gig, or modifying a recording where the instrument you want to change is one of many sounds. In the clean setting (or with electronic instruments in a gig - where technologies such as MIDI sensors are a thing for many instruments) you may be able to ask what you want in dsp.stackexchange.com $\endgroup$ – Neil Slater Nov 8 at 8:53
  • $\begingroup$ @NeilSlater as an exercise in studying AI/ML, it would be as controlled as possible with both instruments electric and using DI interfaces. $\endgroup$ – gator Nov 8 at 15:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.