"MIDI Music Group" by Ferran Carrascosa, Oscar Casanovas, Alessandro Pintaudi and Martí Pomés. Advisor: Carlos Segura. July 2019.
Github link: https://github.com/alepintaudi/music-generation.git
This project presents the implementation of a basic MIDI compositions music generation system while we learn about the implementation of 3 Neural Network models (LSTM Seq2One with Keras, LSTM Seq2One with PyTorch, LSTM Seq2Seq with PyTorch). We also provide an overview of the current commercial and research projects on Music Generation systems using Deep Learning as well as other simpler but very interesting individual projects developed recently.
Over the past 4 years, we have seen impressive progress in the field of generative music and Artificial Intelligence thanks to the developments made on Deep Learning technologies.
The goal of this Postgraduate project is to get hands-on experience on building our own Model that is trained with a collection of MIDI files and then is modified to generate new short composition snippets that are evaluated on their degree of musicality and closeness to passing a Turing test.
It is important to note that none of the four project participants had previous experience implementing Deep Learning models and that all the knowledge has been acquired during the past 5 months over the UPC School - Artificial Intelligence with Deep Learning Postgraduate Degree and also through online resources and publications. We also want to thank our supervisor Carlos Segura Perales for his dedication, availability and valuable insights provided over the past 2 months.
Recent projects such as Bachbot (Feynman Liang), Coconet (Google Magenta), Music Transformer (Google Magenta) and MuseNet (OpenAI) among others, have demonstrated technology capable of achieving musical composition results able to pass the Turing test on certain cases and music genres/instrumentation.
At the same time, several private companies have been working on research and commercial applications of similar systems for certain "mainstream" use cases such as described in the Benchmark Section.
During the latest AI Music news boom, many journalists predict that soon we will see more and more of these generative systems deployed commercially for several real world applications, such as royalty-free, low-cost, quickly personalized and on-demand custom-tailored music.
These use cases could have a strong “Product/market fit” for new automatic music composing systems able to provide professional quality music at a very affordable rate for media companies, freelancers, individuals, music enthusiasts as well as general music listeners.
For reference, we believe it is also important to mention the following Deep Learning Music Generation projects that show the output quality of current State of the art commercial and research applications: