| dc.description.abstract | A typical musical piece often has noticeable recurring segments of music, often referred to as musical patterns, that provide important insights into the structure of the track. There is no established set of rules that define a musical pattern, and the process of human pattern annotation is long, tedious, and highly subjective. Algorithmic versions of such annotation techniques using a structured approach do exist, but perform markedly worse in tasks such as classification when compared to their human counterparts.
We discuss a number of existing pattern discovery algorithms and propose an experimentation framework that utilises long-term memory recurrent neural networks (LSTMs) to attempt to teach a model about the characteristics of individual pattern classes in order to allow that system to identify the same pattern classes in an unstructured environment. The framework consists of three stages, whereby we progressively increase the complexity of musical training data to track and analyse the performance of the proposed system. We also show a preliminary process of parameter tuning that provides a more optimal parameter combination for ensuing training and testing.
Our results show that it is possible for an LSTM-based neural network to establish a definition for pattern classes given enough training time and training data. We have also learned that it is possible to train an LSTM neural network on patterns combined synthetically in sequence that are taken from a realistic dataset and have the learned model be able to identify those patterns in their original scenarios. However, we conclude that the proposed system is still a very basic approach that cannot yet compete alongside current state-of-the-art pattern discovery algorithms, and that the results we provide can serve as a promising baseline for future work on this topic.
GitHub Link: https://github.com/ErikScerri/project-infomgmt-2019 | |
