Bayesian inference model for applications of time-varying acoustic system identification

A major challenge in acoustic signal processing lies in the uncertainty regarding the current state of the acoustic environment. The relevant applications in the field of speech and audio signal processing include the multichannel sound capture, the signal processing for spatial sound control, and the acoustic echo/interference cancellation. In this paper, a Bayesian impulse response model is proposed for acoustic system identification. It is justified by the stochastic nature of time-varying and noisy environments. In particular, we argue for a state-space dynamical model of the unknown impulse responses as a suitable form to incorporate a priori information of the acoustic environment. For the echo/interference cancellation case, we then describe the Bayesian inference of the acoustic system. It is structurally and experimentally compared to maximum-likelihood and least-squares estimators which are both rooted in deterministic system modeling. Algorithmic structure and performance, both speak for the Bayesian inference.