Use of Deep Learning methods such as LSTM and GRU in polyphonic music generation

Music is an essential part of everyone’s life and plays a very important role in many of the media and entertainment industries such as movies, games, television etc. These fields, especially music industry, have an extensive need of an integration of technology and a system that can assist artists in creating better music with ease. This is where the long-term structure creating capabilities of LSTM and GRU be used to create a polyphonic musical piece which can be used to come up with unique ideas for songs by musicians. It can also be used by non-musicians in case they want to create something personalised but do not have the right tools or the underlying theory knowledge of music to create. In this study, two separate neural network models created with LSTM and GRU respectively, are trained on music files and made to come up with patterns based on that pattern knowledge. Those patterns were then evaluated on parameters such as how close to a human composition can the networks predict notes and number of other conditions such as creating repetitive pattern, dissonant notes etc. The study concluded that most people could identify which is actual human composition and which is a machine generated composition, and rated the machine generated compositions at around 70% likeable. The LSTM model was able to learn the song structure such as chorus and verses, and was able to recreate those in its predictions, and GRU had more of repetitive and dissonant notes in the composition