Music for Self-Attention
Abstract
Creating an artificial intelligence aid for music composers requires a practical and modular approach that allows the composer to manipulate the technology as needed in the search for new sounds and ideas. Many existing approaches fail to capture the interest of composers as they are limited beyond their demonstrative purposes or allow minimal interaction with the composer. Score-Transformer (ST) demonstrates a practical integration of artificial intelligence to aid in the creation of new music by working seamlessly alongside any popular notation software. Furthermore, ST can be trained by the user with additional works (including their own compositions), fine-tuning it and minimizing the risk of the software becoming outdated or impractical for continued use. ST was used in the creation of my dissertation piece, Music for Self-Attention.
Music for Self-Attention features an innovative algorithmic approach to traditional compositional methods by demonstrating the benefits of using deep learning in music composition to aid with certain pitch and rhythmic decisions. These artificially generated decisions were not intended to fully remove the human element from composing but rather to work in tandem with the composer, in this case, myself. The piece lasts approximately 23 minutes and loosely follows a form of theme and variations in reverse. This steady process gradually deconstructs a series of variations which were each was initially generated by artificial means until the final movement—containing no artificial intervention—is revealed. This is not a theme and variations in the usual way as the listener won’t hear the theme in each variation. Each movement reflects different processes involved in the training and deployment of artificially intelligent software. Music for Self-Attention is meant to demonstrate a symbiotic relationship that can exist between artificial and human creativity in music composition.