An Exploration of Using Artificial Intelligence Techniques to Learn and Improve Automated Composition Systems

The deep learning model LSTM (Long Short - Term Memory) can be used to deal with the problem of sequences, and is currently the most commonly used model to deal with music training. The process of LSTM training is sequence-to-sequence, which allows the back-and-forth relationship of notes in an instrument to be learnt by the model; however, when dealing with multiple instruments, there are multiple sequences that need to be processed at the same time, which is a limitation of the LSTM model architecture, which makes most of the current automated composition systems for single instrument training and output. However, when dealing with multiple instruments, there are multiple sequences that need to be processed at the same time, which is a limitation of the LSTM model architecture, and this makes most of the current automated composers train and output for a single instrument. Currently it is visible to deal with LSTM multiple sequences problem. Whereas this study needs to achieve multiple sound event sequence inputs corresponding to multiple sequence outputs, and the sequences are connected to each other, a problem for which there is no effective solution at present. This study attempts to choose to use Deep Improvisation (Tatsuya, 2017) for improvement by training a single track and optimizing and modifying the system to be able to have two tracks of input and two tracks of output to achieve a more resilient system, which in turn creates more realistic music. Due to the above background, the motivation to improve the deep learning automatic songwriting system arises with the expectation that the research in this paper will improve the system to be able to produce multi-tracked and more realistic music, so that more researchers will be able to apply the system to be able to produce songs and apply them to a variety of contexts at a small cost.