Statistical Analysis of Conductor-Musician Interaction : With a Focus on Musical Expression

Sammanfattning: The kind of interaction occurring between a conductor and musicians while performing a musical piece together is an unique instance of human non-verbal communication. This Musical Production Process (MPP) thus provides an interesting area of research, both from a communication perspective and by its own right. The long term goal of this project is to model the MPP with machine learning methods, for which large amounts of data are required. Since the amount of data available to this master thesis stems from a single recording session (collected at KTH May 2014) a direct modeling of the MPP is unfeasible. As such the thesis can instead be considered as a pilot project which examines pre-requisites for modeling of the MPP. The main aim of the thesis is to investigate how musical expression can be captured in the modeling of the MPP. Two experiments, as well as a theoretical investigation of the MPP, are performed to this end. The first experiment consists of a HMM classification of sound represented by expressive tone parameters extracted by the CUEX algorithm, and labeled by four emotions. This experiment complements the previous classification of conducting gesture in GP-LVM representation performed by Kelly Karipidou on the same data set. The result of the classification implicitly proves that expression has been transferred from conductor to musicians. As the first experiment considers expression over the musical piece as a whole, the second experiment investigates the transfer of expression from conductor to musician on a local level. To this end local representations of the sound and conducting gesture are extracted, the separability of the four emotions are calculated for both representations by use of the Bhattacharyya distance and the results are compared in search for correlation. Some indications of correlation between the representations of sound and gesture are found. The conclusion is nevertheless that the utilized representations of conducting gesture do not capture musical expression to a sufficient extent.