Symbolic piano music understanding from large-scale pre-training

Abstract

Pre-training driven by a vast amount of data has shown great power in natural language understanding. The existing works using pretraining for symbolic music are not general enough to tackle all the tasks in musical information retrieval. To make up for the insufficiency and compare it with the existing works, we employed a BERT-like masked language pre-training approach to train a stacked Music Transformer on polyphonic piano MIDI files from the MAESTRO dataset. Then we finetuned our pre-trained model on several symbolic music understanding tasks. In our current work in progress, we complemented several note-level tasks, including next token prediction, melody extraction, velocity prediction, and chord recognition. And we compared our model with the previous works.