Coherent Music Composition with Efficient Deep Learning Architectures and Processes

In recent years, significant advancements in music-generating deep learning models and neural networks have revolutionized the process of composing harmonically-sounding music. One notable innovation is the Music Transformer, a neural network that utilizes context generation and relationship tracking in sequential input. By leveraging transformer-based frameworks designed for handling sequential tasks and long-range functions, the Music Transformer captures self-reference through attention and excels at finding continuations of musical themes during training. This attention-based model offers the advantage of being easily trainable and capable of generating musical performances with long-term structure, as demonstrated by Google Brain’s implementation. In this study, I will explore various instances and applications of the Music Transformer, highlighting its ability to efficiently generate symbolic musical structures. Additionally, I will delve into another state-of-the-art model called TonicNet, featuring a layered architecture combining GRU and self-attention mechanisms. TonicNet exhibits particular strength in generating music with enhanced long-term structure, as evidenced by its superior performance in both objective metrics and subjective evaluations. To further improve TonicNet, I will evaluate its performance using the same metrics and propose modifications to its hyperparameters, architecture, and dataset.