Growing up I was always interested in electronic music, dj mixes, electronic music composition etc. I used to make my own hobby mixes as well.
When making mixes, DJs traditionally use a mixer to mix elements of two tracks to make somewhat seamless and exciting transition from one track to another. A two channel mixer usually has three knobs to filter out high, middle and low sounds for each track.
The low knob is supposed to filter out the drums and kicks.
The middle knob is supposed to filter vocals and melodies
The high know filters out snares etc, the high sounds.
But as you can imagine this kind of filtering is not enough to create clean mixes because instruments, sounds, vocals can span across wide frequencies. I think every dj wishes to have a tool that could separate music tracks into its components (vocals, instruments, drums, kicks) cleanly.
In my wild dreams, I would like to make an app that could separate a track into its components in real time, so that a user could pick and choose what ever he/she wants to include in a mix or other purpose. I have no idea if this is practically feasible. But learning a little about signal processing , Fourier transform and deep learning gives me some hope and much excitement. I am not expert in any of this fields. So I will have to learn a great deal about each topic and find out what is already have been achieved. I imagine I am not the only one come up with this idea. But I didn't come across a popular app that does what I want.
Heilmeier Catheism Checklist
Is it doable?
I believe it is but I have to make a literature review to see what is already out there and what can I improve
Is it revolutionary?
If I can come up with an app that does this in real time it would certainly be revolutionary.
Is it feasible?
That I have to find out
Is it interesting?
It is very interesting.
Is it an area You want to learn about?
In order to realize this project I will have to learn a lot about signal processing, audio programming, machine learning, deep learning, app development etc. Even if I fail, I can use my new found knowledge in other areas.
One thing that is interesting about this is that you can probably find a *very* large amount of labelled training data, or you could make it yourself. If you can find music (or make music) that comes from a drum track and a singing track, etc... you can create a dataset that has the full music (the two tracks combined), and then the two separate tracks (which is the desired output).
Crazy plan of this could be to have a video of a band that is playing a song, and you can click on, say, the drummer, to select the audio that you want to modify. This YouTube video might give some inspiration: https://www.youtube.com/watch?v=rVQVAPiJWKU