Generative Adversarial Neural Networks for Random and Complex Chord Progression Generation

Abstract

In the latter years, recurrent neural networks for generative modelling of sequences with long range dependencies have been outperformed by the use of autoregressive models such as Transformer XL. With these models, a complete processing of long range dependencies has been increasingly improved and optimized in terms of memory and execution time. Our experiments use Google BERT, a transformer-based machine learning technique for natural language processing that can produce high fidelity text generation outputs. We use this technique to work on music, as this is represented by symbols and can be considered a type of language. Jazz oriented or influenced music composition often includes dealing with harmonic progressions in which each chord usually consists of four or more different notes. To design complex chord progressions is a very specialized task and expensive in time that demands lots of hours of music theory studying. Furthermore, currently there aren't apps and music industry plugins designed to generate this type of output data interactively. To work on this we trained a generative adversarial neural network that operates using two Transformer XL as classifier and generator respectively. We trained the model on a self designed corpus of MIDI files that explicitly considers the harmonic patterns of various well known Jazz themes, leaving behind lead melody and rhythm patterns. After achieving an stable training of the model we were able to generate extensive batches of complex harmonic progressions on MIDI files. These outputs clearly expose stylistic properties that are present in Jazz music chord progressions.