Download available

Authors:

Publication or Conference Title:

Abstract:

Two types of artificial neural networks are used to determine a sufficient subset of data for reasonable classification of musical instrument based on performance data from motion capture. Feedforward and recurrent networks are trained on subsets of joint angles and centre of mass from performances by violists and clarinettists. A successfully learned mapping from the reduced set of input data to the correct instrument classification implies that the data subset carries sufficient information to identify an instrument. After training, cross- validation is performed by testing networks on previously unseen data. Finally, performance is compared with that of humans performing similar classification tasks based on the same data.
Feedforward and recurrent networks are found to perform similarly when classifying test data. Instrument recognition rates by networks are comparable with human recognition rates over the various data subset conditions. The methods demonstrated here could also be applied to other movement analysis domains (e.g. gait, posture, kinematics, clinical/rehabilitation work).

Publication Details: