Investigating representation of tablature data for NLP music prediction

Abstract

In this thesis, the ability of CharRNN models learning to compose guitar music using varying representations of guitar tablature is explored. I utilize a well-versed sequential model of LSTM cells, and investigate the ability of said model to input, and predict both character to character, and sequence to sequence, following the principles of natural language processing and music information retrieval. The study was conducted on datasets consisting of data naïvely retrieved from a subset of classical guitar tablature. With regards to tablature structure, the experiments uncover a clearly superior form for character to character prediction, producing a model capable of composing seemingly musically coherent phrases. The work is not fully able to compare the character predictor with the sequence predictor and further details how this could potentially be alleviated.