Research visit to Singapore at A*STAR and IEEE TENCON
Last week I was in Singapore to present the MorpheuS research I have performed on my Marie-Curie grant at IEEE TENCON. The title of my talk was "MorpheuS: Automatic music generation with recurrent pattern constraints and tension profiles". I was also invited to give a seminar on "Machine learning and optimization applied to digital music" at the High Performance Computing Institute at A*STAR in Singapore. A short abstract of the talk at A*STAR:
In this talk, we will explore how state-of- the-art machine learning and optimization techniques are currently being applied to further the field of digital music. Research projects that will discussed include automatic music generation, dance hit prediction and automatic piano fingering.
Music generation systems have attracted research attention since the advent of computing. They have become increasingly important, bolstered by rising global expenditure on digital music, which was over 64 billion USD in 2014 alone. Most music generation systems are based on statistical models and rules. A drawback of these early systems is their inability to synthesize music that possesses global structure. When music does not have a clear direction or long-term coherence, it fails to hold the listener’s attention and can be hard to follow. The problem of structure has recently been tackled using deep learning, with mixed results. Guaranteed success has been achieved using optimisation algorithms that constrain the structure of the generated music by incorporating pattern detection techniques. This approach is currently being examined in the authors’ EU project MorpheuS. This project also implements a model for calculating tonal tension that allows the user to generate polyphonic piano music according to a given tension profile.
In the second part of this talk, we will zoom in on the dance hit song prediction problem. With annual investments of several billions of dollars worldwide, record companies can benefit tremendously by gaining insight into what actually makes a hit song. We will describe how we built a database of dance hit songs from 1985 Until 2013. Different classifiers: such as SVM and logistic regression are then used to build and test prediction models. The resulting model has a good performance when predicting if a song is a dance hit versus a lower position listed. Thirdly, we will also discuss how optimization algorithm can be used to calculate the optimal fingering of complex polyphonic piano pieces.