Audacity has added AI audio editing capabilities thanks to Intel’s free OpenVINO plugins. These plugins add AI-powered noise suppression, speech transcription, music generation and remixing, and music separation to the freeware sound editor and are available for download today.
AI models are often multiple gigabytes, tbh it’s a good sign that it’s not “AI” marketing bullshit (less of a risk with open source projects anyway). I’m pretty wary of “AI” audio software that’s only a few megabytes.
Why are they that big? Is it more than code? How could you get to gigabytes of code?
They’re composed of many big matrices, which scale quadratically in size. A 32x32 matrix is 4x the size of a 16x16 matrix.
Currently, AI means Artificial Neural Network (ANN). That’s only one specific approach. What ANN boils down to is one huge system of equations.
The file stores the parameters of these equations. It’s what’s called a matrix in math. A parameter is simply a number by which something is multiplied. Colloquially, such a file of parameters is called an AI model.
2 GB is probably an AI model with 1 billion parameters with 16 bit precision. Precision is how many digits you have. The more digits you have, the more precise you can give a value.
When people talk about training an AI, they mean finding the right parameters, so that the equations compute the right thing. The bigger the model, the smarter it can be.
Does that answer the question? It’s probably missing a lot.
The current wave of AI is around Large Language Models or LLMs. These are basically the result of a metric fuckton of calculation results generated from running a load of input data in, in different ways. Given these are often the result of things like text, pictures or audio that have been distilled down into numbers, you can imagine we’re talking a lot of data.
(This is massively simplified, by someone who doesn’t entirely understand it themselves)
And this is why current AI is nowhere near sentience or anything else that occasionally comes up in the press. Ultimately it’s an algorithm (or set of algorithms) which ingest a bunch of training data and later regurgitate the most common patterns, albeit with some context-sensitive cues thrown in.
That’s one reason why AI image generators can’t, on average, produce images of e.g. a broken radio: because there are no broken radios in their training data.
It’s data
It’s really nothing of the sort.
There are graph neural networks (meaning NNs that work on graphs), but I don’t think that’s what is used here.
I do not understand what you mean by “routes”. I suspect that you have misunderstood something fundamental.
You can see a neural net as a graph in that the neurons are connected nodes. I don’t believe that graph theory is very helpful, though. The weights are parameters in a system of linear equations; the numbers in a matrix/tensor. That’s not how the term is used in graph theory, AFAIK.
If you look at the nodes which are most likely to trigger from given inputs then you can draw paths
Tensorflowlite models are tiny, but they’re potentially as much an audio revolution as synthetizer were in the 70s. It’s hard to tell if that’s what we’re looking at here.