Deep Plunder (2020)

David Braun

Abstract

We use a deep recurrent neural network to identify which regions of music recordings contain isolated music production samples. We use parallel multiprocessing to generate 50 hours of fake music consisting of randomly chosen library samples and real song excerpts sequenced with various augmentations. The neural network uses supervised learning with an audio spectrogram as input and a one-dimensional labeled region as output. We achieve low error rates on our synthesized train and test datasets but inadequate qualitative results on real music. We believe our network has potential to generalize to real music, but to avoid time-consuming manual labeling, we must research better data generation and augmentation techniques.

Full paper: Deep Plunder - Braun (2020).pdf

Example audio file: plunder_0017-(8kHz).wav

Starter code for piano transcription (not Deep Plunder): Onsets-Frames