When I was in a band, we used a simple digital audio recorder to document our sessions. The annotations look helpful for trying to remember what you were working on in the previous session. How did you perform the instrument detection?
We collected existing recording data from musicians (that website is still online: https://tape.it/needsyourhelp), annotated it manually with Prodi.gy, and then trained a YAMnet model for the classification.
Granted, had we known that Apple would ship a much improved version of their own sound classifier in iOS 15, we probably wouldn't have invested so much work in it - their new pretrained model is truly outstanding (although it only works on iOS >= 15). But I don't regret doing it; we learned a lot about mobile deployment of ML models, which will come in very handy when we deploy our next ML features for audio quality improvements.
