Music generation using tracker music and machine learning

Sammanfattning: We investigate the modelling of polyphonic “tracker music” using deep neural networks. Tracker music is a music storage format invented in the late 1980s for use on that time’s home computers and is often used for storing synthesized electronic music. Tracker music differs significantly from other music formats and has properties that makes it both harder and easier to use for training neural networks than other formats. This makes it interesting to explore what methods are most suitable for extracting musical information from the format. As far as we know, we are the first to explore how to use tracker music for music generation. We design a method for turning tracker music into sequential data usable for training neural networks. The sequential nature of the data means that musically unaware sequence models can be used for training. The method is general and can be applied to other kinds of symbolic music. We then compile a dataset of about 20 000 freely available instrumental songs in the tracker format MOD, downloaded from the website The Mod Archive. We use the dataset to train several different sequence models, including a Long Short-Term Memory (LSTM) network and a Transformer model. We evaluate the models using a sequence completion task and we investigate the statistical properties of the output. We also conduct a listener study involving some 100 participants to determine how often music generated by the models is preferred over human-composed music. The listener study’s result indicates that music generated by the models trained on the dataset is sometimes competitive with music composed by humans. We conclude that neural networks for music generation can be trained using tracker music using our proposed conversion method, but that it is cumbersone. Due to how the tracker music format is constructed it is significantly more difficult to get musical information out of it than we initially thought.