Modelling Gestures in Music Performance with Statistical Latent-State Models

We discuss try to identify ”gestures” in music performances by observing patterns in both compositional and expressive properties, and by modelling them with a statistical approach. Assuming a finite number of latent states on each property value, we can describe those gestures with statistical latent state models, and train them by unsupervised learning algorithms. Results for several recorded performances indicate that the trained models could identify the gestures observed, and detect their boundaries. An entropybased measure was used to estimate the relevance of each property for the identified gestures. Results for a larger corpus of recorded and annotated musical performances are promising and reveal potential for further improvements.