Meter Detection and Alignment of MIDI Performance

Metrical alignment is an integral part of any complete automatic music transcription (AMT) system. In this paper, we present an HMM for both detecting the metrical structure of given live performance MIDI data, and aligning that structure with the underlying notes. The model takes as input only a list of the notes present in a performance, and labels bars, beats, and sub beats in time. We also present an incremental algorithm which can perform inference on the model efficiently using a modified Viterbi search. We propose a new metric designed for the task, and using it, we show that our model achieves state-of-the-art performance on a corpus of metronomically aligned MIDI data, as well as a second corpus of live performance MIDI data. The code for the model described in this paper is available at https://www.github.com/apmcleod/met-align.