Machine learning of Finite-Difference Time Domain (FDTD) physical modelling sound simulations of drumhead paste pattern distributions

Abstract

Drummers attach different kinds of material on their drumheads to either increase damping or to tune them and adjust the relationships of sound partials. The former is a common practice for drummers, while the latter may be found in percussion instruments of various ethnic traditions, such as the Myanmar pat wain drum circle or the Indian tabla. A Finite-Difference Time Domain (FDTD) physical model of a drumhead was used to compute more than 2000 sounds simulating membrane vibration, which was adjusted by adding varied amounts of paste, distributed in different surface patterns. These sounds were analysed using Self-Organizing Maps (SOMs) as well as a Convolutional Neural Network (CNN). The SOMs were used to cluster the partial relationships of the generated sounds. It is demonstrated that different paste patterns correspond to different clusters. Furthermore, the CNN was trained to identify the damping approach, yielding an accuracy of 94% for paste pattern classification and a mean error of +/-11% for the estimation of membrane mass increase. These tools can be used to identify damping patterns used in historical drum recordings or as suggestions to percussionists for deriving a desired sound texture.