Where I try to teach a computer to know about electronic music. This is very much a work-in-progress.
(If it seems like I don't know exactly what I'm doing, you are entirely correct. Please help!)
Mostly a series of Python sketches in TensorFlow and Keras trained via with PyRekordBox which creates a robust pre-labeled dataset of music with labels from RekordBox's phrase analysis system.
My end goal is to have "phrases" in electronic music mostly properly identified in real-time so I can make lighting (like in WLED MoonModules react to the musical "movement" and not just the waveforms.
Mostly tested and trained for what I play when I DJ (House-related, NuDisco, house remixes of older songs, etc in a 120-130ish BPM range) - which is what RekordBox generally calls the "high" mood, and 5 tags of 'high_intro', 'high_chorus', 'high_down', 'high_up', and ,'high_outro'.
The system captures a log of the files as it suimulates playback, and saves some of the mislabeled segments for later retraining. A bad result later is actually a good training result, as the system will learn to generalize better - I think.
The current model seems to generalize pretty well - I have tested a few songs at every possible 3-second sample in the song (across 3000 offsets) and the analysis is pretty good at at recognizing a label even when alignment varies.
Training testing continues, and I have some ideas for scanning a file more quickly than taking 3000*seconds_of_length/3 samples to figure out how well it's generalizing - because it seems best to train on the "worst" offsets rather than random offsets. With the current results, I'm watching more for an overall improvement in the correct % scores rather than a high score - unless every file suddenly jumps together.
I will repeat that I'm basically making very vaguely educated guesses at this point - but I wanted to release something with the hope that prehaps more eyeballs will make it better.
A huge thanks to Dylan Jones for PyRekordBox upon which all the labeling is derived.