Algorithmic Ability to Predict the Musical Future: Datasets and Evaluation

Music prediction and generation have been of recurring interest in the field of music informatics: many models that emulate listeners' musical expectancies, or that produce novel musical content have been introduced over the past few decades. So far, these models have mostly been evaluated in isolation, following diverse evaluation strategies. Our paper provides an overview of the new MIREX task Patterns for Prediction. We introduce a dataset, which contains monophonic and polyphonic data, both in symbolic and audio representations. We suggest a standardized evaluation procedure to compare algorithmic musical predictions. We compare two neural network models to a baseline model and show that algorithmic approaches can correctly predict about a third of a monophonic segment, and around half of a polyphonic segment, with one of the neural network models achieving best results. However, other approaches to algorithmic music prediction are needed to achieve a more rounded picture of the potential of state-of-the-art methods of music prediction.