Control parameters inversion using genetic algorithms applied to numerical impedance synthesis for woodwinds
Laura Perichon, Olivier Carrière, Jean-Pierre Hermand, Matthias Meyer, Philippe Guillemain · The Journal of the Acoustical Society of America · 2006
Woodwind sound synthesis in real time has been achieved using different computational techniques, especially the so-called numerical impedance model based only on acoustical variables. The impedances of the excitation mechanism and of the resonator are approximated by IIR filters and related by a nonlinear coupling. The control parameters can be reduced to a number of three including the length of the resonator, the mouth pressure, and a mask parameter. In this paper we present preliminary inversion results for the latter two parameters in the case of a clarinet. The inversion algorithm is based on differential evolution metaheuristics with a number of prior conditions on the correlated parameters. A cost function based on time-domain correlation has been implemented, allowing us to obtain better results in terms of signal matching than with conventional descriptors extraction. The technique has proved its effectiveness on the sustained part of synthesized sounds and a regularization method is presented to allow us to invert attack and decay parts as well as real analogically recorded signals. Results will be of main interest for tuning control parameters in real-time model-based synthesis.