Towards Playing in the 'Air': Modeling Motion-Sound Energy Relationships in Electric Guitar Performance Using Deep Neural Networks

Abstract

In acoustic instruments, sound production relies on the interaction between physical objects. Digital musical instruments, on the other hand, are based on arbitrarily designed action--sound mappings. This paper describes the ongoing exploration of an empirically-based approach for simulating guitar playing technique when designing the mappings of 'air instrument' designs. We present results from an experiment in which 33 electric guitarists performed a set of basic sound-producing actions: impulsive, sustained, and iterative. The dataset consists of bioelectric muscle signals, motion capture, video, and audio recordings. This multimodal dataset was used to train a long short-term memory network (LSTM) with a few hidden layers and relatively short training duration. We show that the network is able to predict audio energy features of free improvisations on the guitar, relying on a dataset of three distinct motion types.