Link for the tool: https://www.melodybot.com/
Composing music automatically is perhaps one of the most crucial, but also difficult, projects in the field of information reconstruction. For expert users it is the most effective means of communication, while for all the others is just one of the best ways of expressing their feelings. However, the composition of new and interesting tracks is a process that requires deep knowledge, experience and expertise. A similar difficulty could also be met in computers, where, despite all efforts, it has been proved to be a particularly demanding task, which has been successful only in a few categories of hearings. As it happens with most of the information retrieval and reconstruction projects in the field of music, systems of automatic composition that have been developed within this dissertation tend to replace the stages of signal processing and extraction from statistical models with architectures of deep learning architectures. For this reason, the traditional way of music representation has been chosen for the purposes of the current dissertation, which is encoding in sequential form and encoding by using the Midi protocol, if we want to be more specific. For the experimental part of the work different neural network architectures were trained so that to create an original sequence of a given initial melody. More specifically, the following were used: a recurrent neural network of deep Long-Short Term Memory (LSTM) with multiple levels, an Encoder-Decoder architecture (LSTM Encoder-Decoder) as well as an Encoder-Decoder architecture with an attention mechanism (LSTM Encoder- Decoder with Attention). Along with the architecture, also the whole training dataset had been changing, using songs from different instruments as: piano, guitar as well as different combinations and variations of them. Finally, some of the hyperparametrs of the above networks were changed, such as LSTM memory size and prediction method, in order to investigate their role and impact on the compositions.