Continuous Time Recurrent Neural Networks for Generative and Interactive Musical Performance

‘Continuous-Time Recurrent Neural Networks for Generative and Interactive Musical Performance’ was a paper delivered at the 4th European Workshop on Evolutionary Music and Art, subsequently published in proceedings.

The paper begins by identifying the dominant paradigm located at the junction of computer music and artificial intelligence as one which identifies the goal of developing generative or interactive software agents which exhibit musicality.

Where once such a goal remained the preserve of dedicated research institutes with access to massively parallel computer networks, the authors demonstrate that this is no longer the case with increased computational capacity in consumer hardware and extensible music programmes such as Max/MSP, PD and Supercollider.

The researchers identify a problem with this potentially liberating convergence of raw number crunching and flexible programming languages: musicians tend to tailor their own dedicated solutions to their own perceived needs. Rather than follow this route, the researchers determined to engineer a generic behavioural tool that can be developed in different directions by different practising musicians, each with their own aesthetic instincts and compositional requirements.

The vehicle through which the researchers develop their solution is a specific form of artificial intelligence called a Continuous-Time Recurrent Neural Network. The researchers evoke a number of previous approaches that have sought to simulate complex systems in both analogue and computer-based settings. The advantage of their chosen model is that the AI is able to evolve in response to stimuli - in this case musical events – and for that evolution to function as a kind of training, meaning that the system can be programmed to adapt to different kinds of musicians’ work. The researchers argue that a CTRNN-based evolutionary music system is ideally suited to working with improvising musicians. They provide evidence for this argument, in the form of applied case studies.

Current Research
I am working with CRiSAP co-director Cathy Lane on a project to develop audio analysis software, in particular for use by composers producing artistic works with collections of speech recordings. This software will enable users to search through a database of sound for similarities in timbre or pitch content. We are making use of a state-of-the-art set of open-source low-level tools currently being developed by the OMRAS II group at Goldsmiths, University of London, which are designed to perform this task over a network for Google-style searching of music databases. Our project involves embedding this existing software in a graphical user interface program which composers and sonic artists can use creatively on their own collections of recorded sound, and aims to make the software accessible to end users without needing an understanding of the low level behaviour of audio similarity technology. As well as the development of the software, the project also involves research into the artistic practice of composing music using automatic analysis audio similarity, and in particular the specific case of the analysis of vocal sounds, and some creative works will be developed as an outcome of the project. In general this project reflects CRiSAP's commitment to research into the development of new software tools for the sonic arts, and their philosophy that the rapid development of prototype software for creative use is a valuable goal.

• Author
• Author