A STATE-SPACE RECURRENT NEURAL NETWORK MODEL FOR DYNAMICAL LOUDSPEAKER SYSTEM IDENTIFICATION

A majority of adaptive systems in audio and acoustic signal processing, for instance in acoustic echo control, relies on finite-impulse response (FIR) modeling for the sake of simplicity and inherent stability. In this paper we consider the problem of dynamical loudspeaker modeling and use a contemporary machine-learning framework for infinite-impulse response (IIR) acoustic system identification. Specifically, we convert a physically-motivated parametric dynamical loudspeaker model to a multivariate state-space representation and solve the parameter identification problem via an equivalent recurrent neural network (RNN). With an appropriate training procedure, we obtain a stable computational system with a correspondence of its few parameters to the physical loudspeaker model. Besides the verification on synthetic data provided by the physical model, we also apply the procedure to a set of real loudspeaker recordings. With appropriate model order, the neural network delivers a simple yet accurate recursive loudspeaker representation.