A Comparative Study of Neural Models for Polyphonic Music Sequence Transduction

Abstract

Automatic transcription of polyphonic music remains a challenging task in the field of Music Information Retrieval. One under-investigated point is the post-processing of time-pitch posteriograms into binary piano rolls. In this study, we investigate this task using a variety of neural network models and training procedures. We introduce an adversarial framework, that we compare against more traditional training losses. We also propose the use of binary neuron outputs and compare them to the usual real-valued outputs in both training frameworks. This allows us to train networks directly using the F-measure as training objective. We evaluate these methods using two kinds of transduction networks and two different multi-pitch detection systems, and compare the results against baseline note-tracking methods on a dataset of classical piano music. Analysis of results indicates that (1) convolutional models improve results over baseline models, but no improvement is reported for recurrent models; (2) supervised losses are superior to adversarial ones; (3) binary neurons do not improve results; (4) cross-entropy loss results in better or equal performance compared to the F-measure loss.