Modeling of Gesture-Sound Relationships in Recorder Playing: A Study of Blowing Pressure

This work deals with instrumental gestures and their relation to sound in recorder playing. In particular, it presents a study of the relationship between the blowing pressure and a number of timbre characteristics of produced sound.

Blowing pressure is acquired by means of a pressure transducer mounted on the mouthpiece of the recorder, while sound is acquired using a microphone and analyzed in the spectral domain via SMS in order to extract spectral shape descriptors corresponding to odd harmonics, even harmonics and residual components. A performance recording is carried out so that a number of representative playing contexts are covered. Collected data (blowing pressure, produced sound, and spectral analyses) are segmented and used to assemble a multi-modal performance database.

Starting from the acquired data (blowing pressure and spectral descriptors), supervised machine learning techniques are used to construct and evaluate two types of models. The first model takes spectral descriptors as input and produces an estimation of the blowing pressure, in a frame-by-frame fashion. The second model takes the blowing pressure as input and produces an estimation of the spectral descriptors, also in a frame-by-frame fashion. While the former lays in the domain of indirect acquisition of instrumental gestures, the application context of the latter corresponds to spectral-domain timbre synthesis. Both models are evaluated via crossvalidation, while the synthesis-oriented model is tested by means of additive synthesis via SMS.