Beat tracking from conducting data with musical pieces

Abstract

Conducting is an act of communication between the conductor and players by human motion in real-time, and beats are the key to analyzing the conducting motion. However, it is hard to follow conducting motions without contextual information because there are different types of conducting styles. In this thesis, we will solve the conducting beat tracking problem using dynamic programming and neural networks. We created ground-truth beat labels to perform a quantitative evaluation and supervised learning, and we suggest two new approaches that use contextual information include the tempo of musical pieces. The handcrafted method detects bouncing motions and fast strokes, and the recurrent neural networks are trained in a supervised manner. Both approaches revealed good performance, in the quantitative and qualitative analysis.