WANNA GO FURTHER

[proud acorn]

hey guys. i work as a sound designer and im currently making models an all is well. but i want to learn deeper. like what kind of learning process and sense is behind this process etc.. thanks in advance

[quartz cloak]

Ayo? @proud acorn level 1 !!!

[trail scroll]

@random grail might be able to help you =3

[random grail]

ask questions in order and all will be answered

[proud acorn]

hey @random grail from here or in private?

[random grail]

doesn't rlly matter
i'd prefer in public so it can be citated by others

[proud acorn]

ok. 1- how does the learning algortihm works? does it learn how any person goes freq to freq in which syllable? 2- if you can help the ai learn a persons voice so it can learn other musical stuff like which drum i like to use or what kind of guitar tone i tend to use by learning my previous projects for example? sorry btw english is not my native language. thanks in advance.

[random grail]

1 - the learning process extracts specific features of the voice and learns how the voice acts depending on the syllable and the sound + dynamics
2 - ai is essentially a prediction algorithm. it learns data it is given and outputs a prediction. you can make any AI, quite literally. if you want, you can make AI predict the temperature outside during the specific day of the winter, you can make ai predict if something is fake, etc etc

[proud acorn]

thanks alex. so if my given data is a logic project for ex. how does it recognize it and the channels inside the project? correct me if im wrong but i feel like at first the ai should know what logic pro is and how to extract and learn the channels one by one. where to start?

[random grail]

well if you want the ai do recognize the logic pro DAW projects and determine what they are about, you first need to find how exactly those projects are encoded. then, preprocess them to extract data like channels, tracks, etc
having all this data you can make the AI deduce (classify) what project is of what genre, for example.
e.g. it sees distortion plugins fx and a few guitars, it deduces it's a metal/rock song
or it sees some violins, flutes, recorders, etc and deduces it's an orchestral song

so you'd have to rip out the useful data from the project files themselves and only then feed it into AI to learn on

[proud acorn]

that was really helpful alex. i dont want to keep bothering you but i have one last question. i can rip out the useful data but i dont know how to feed it to ai. i think this is the part that my audio knowledge becomes worthless 😄

[random grail]

you just grab the data and encode it slice by slice for machine to understand and learn from
in other words, split dataset into equal-length chunks and feed them to the model as an input, then use the backwards pass and optimizer to lower the loss over time

you'd have to design an architecture that is capable of doing exactly that, then finetune value one by one to give the best results
tbh there are some advice about how to approach these, but most of the times you basically make educated guesses and use arbitrary numbers that you then adjust to see what happens to the result

[proud acorn]

i will look into that. thank you alex.