MULTI-TASK LEARNING FOR BEAT DETECTION ON MIDI PERFORMANCES

Abstract

Beat detection is a fundamental task in music information retrieval as it necessarily requires understanding of musical structure that is essential for other tasks such as music transcription and harmonic analysis. In this thesis, we propose a transformer architecture within a multi-task framework to learn to predict beat positions in musical sequences using symbolic representations. We train our model on a dataset consisting of performances of classical piano music, where beat annotations have been provided by musical professionals. Based on principles of transfer and representation learning, we select a collection of sequence to sequence tasks that benefit from having shared weights in the encoders and decoders of the model. We show the effectiveness of including additional training tasks, and we provide both quantitative evaluation (using F-measures) and qualitative evaluation (analysis of examples). We also discuss the challenges of evaluating the beat detection task.