Physical Modeling meets Machine Learning: Teaching Bow Control to a Virtual Violinist

The control of musical instrument physical models is difficult; it takes many years for professional musicians to learn their craft. We perform intelligent control of a violin physical model by analyzing the audio output and adjusting the physical inputs to the system using trained Support Vector Machines (SVM). Vivi, the virtual violinist is a computer program which can perform music notation with the same skill as a beginning violin student. After only four hours of interactive training, Vivi can play all of Suzuki violin volume 1 with quality that is comparable to a human student. Although physical constants are used to generate audio with the model, the control loop takes a “black-box” approach to the system. The controller generates the finger position, bow-bridge distance, bow velocity, and bow force without knowing those physical constants. This method can therefore be used with other bowed-string physical models and even musical robots.