Guitar Lessons by Chip McDonald - Artificial Intelligence / Machine Learning in Guitar Sims is NOT "Amp Moddeling"

Monday, August 8, 2022

Artificial Intelligence / Machine Learning in Guitar Sims is NOT "Amp Moddeling"

Suddenly, across multiple musician-based forums, as well as social media I've seen people now acknowledging that, yes, machine learning is going to be the Next Big Thing in music. There is still much cognitive dissonance. I see people referencing the Kemper's process of using test tones as being related to "artificial intelligence" and "machine learning". I have also seen people reference circuit modeling as being the same, and again - it is not.
Having the word "Neural" in a companie's name also doesn't mean they're automatically doing something a.i./ML related. On the otherhand, I know that the company IK Multimedia IS claiming to be coming out with a machine learning based guitar sim system soon, that I eagerly await to see how it works. It's not the first ML based guitar sim. On github there are a few, but they're fairly basic (although very useable if one is technically savvy). IK Multimedia though, claims to have put together a package that allows the user to use their own gear (or others) to train the dataset. They also claim what amounts to a hard to believe quick time in doing it as well. Machine learning as a process is "sort of" a way to make the numbers of an input dataset be arranged so that they have characteristics of another dataset. This is commonly shown in visual examples, and sometimes in audio where a person is made to speak and sound like someone else. This is accomplished by "training" the ML process on a dataset of examples you want the output to resemble. This requires a LOT of computing horsepower; high end graphics video cards, and hours and hours at a minimum. The longer you let the process cogitate over the dataset, the better/more realistic the output. As well as how currated the training data is. IK Multimedia claims to have this down to 15 minutes. I'm fairly sure they're taking some shortcuts, and it will probably require the user to upload a dataset to a website, where they might have a "farm" of graphic cards to use to train the dataset. However, I'm also fairly sure it will probably create results that will get guitar simulation out of the proverbial Uncanny Valley. The question will be, will they be able to get this out to the public before some other upstart? I've experimented with PyTorch, and some of the ML things on GitHub but quickly realized I need the aforementioned graphics card - and a lot more time and patience with Java, Python to do what I think needs to be done. It wouldn't be much for most competent programmers to put together a ML based guitar sim package; but to make it work for everyone is the trick. I'm very curious to see what IK Multimedia has come up with...

No comments:

Post a Comment