Beatty: Automatic Tempo Curve Synthesis for Expressive MIDI Track Playback

Abstract

Beatty is a sequence-to-sequence machine learning model to predict expressive timing decisions for excerpts of classical solo piano music. Composed of a bidirectional encoder LSTM and decoder LSTM with attention, Beatty predicts tempo labels based on input note sequences. The input note sequence is obtained by transforming a MIDI file representation of the musical score into a series of one-hot note vectors, which encode the MIDI note pitches, velocities, and durations, and are augmented with additional harmonic tension information. The target output is a sequence of tempo labels, represented as ratios of the sequence’s initial starting tempo. We demonstrate that the harmonic tension augmentation, as well as learning from filtered tempo label sequences, improve model performance. In qualitative evaluation, the model output receives positive feedback when its predicted tempo sequence is subtle and smooth and criticism when it fluctuates too greatly, suggesting areas for future exploration and improvement.