Master of Science (MS)
The recent explosion of interest in deep neural networks has affected and in some cases reinvigorated work in fields as diverse as natural language processing, image recognition, speech recognition and many more. For sequence learning tasks, recurrent neural networks and in particular LSTM-based networks have shown promising results. Recently there has been interest – for example in the research by Google’s Magenta team – in applying so-called “language modeling” recurrent neural networks to musical tasks, including for the automatic generation of original music. In this work we demonstrate our own LSTM-based music language modeling recurrent network. We show that it is able to learn musical features from a MIDI dataset and generate output that is musically interesting while demonstrating features of melody, harmony and rhythm. We source our dataset from VGMusic.com, a collection of user-submitted MIDI transcriptions of video game songs, and attempt to generate output which emulates this kind of music.
Mauthes, Nicolas, "VGM-RNN: Recurrent Neural Networks for Video Game Music Generation" (2018). Master's Projects. 595.