Predicting Marimba Stickings Using Long Short-Term Memory Neural Networks

In marimba music, `stickings' are the choices of mallets used to strike each note. Stickings significantly influence both the physical facility and expressive quality of the music performance. Choosing 'good' stickings and evaluating one's stickings are complex choices, often relying vaguely on trial-and-error. Machine learning (ML) approaches, particularly with advances in sequence-to-sequence techniques, have proved suited for similar complex classification problems, motivating their application in our study. We address the sticking problem by developing Long Short-Term Memory (LSTM) models to generate stickings in 4-mallet marimba music trained on exercises from Leigh Howard Stevens' Method of Movement for Marimba. Model performance was measured under a range of metrics to account for multiple sticking possibilities, with LSTM models achieving a maximum average micro-accuracy of 97.3%. Finally, we discuss qualitative observations in sticking predictions and limitations of this study and provide direction for further development in this field.