Real-time bowed string feature extraction for performance applications

Although the physics of the bowed violin string are well understood, most audio feature extraction algorithms for violin still rely on general-purpose signal processing methods with latencies and accuracy rates that are unsuitable for real-time professional-calibre performance. Starting from a pickup which cleanly captures the motion of the bowed string with minimal colouration from the bridge and body, we present a lightweight time-domain method for modelling string motion using segmented linear regression. The algorithm leverages knowledge of the patterns of Helmholtz motion to produce a set of features which can be used for control of real-time synthesis processes. The goal of the paper is not a back-extraction of physical ground truth, but a responsive, low-latency feature space suitable for performance applications