We collected existing recording data from musicians (that website is still online: https://tape.it/needsyourhelp https://tape.it/needsyourhelp), annotated it ma

We collected existing recording data from musicians (that website is still online: https://tape.it/needsyourhelp 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.