Towards optimal regularization by incorporating prior knowledge in an acoustic echo canceller

The problem of poor excitation is often encountered in acoustic echo cancellation, due to the high coloration of audio signals and the large dimension of the room impulse response parameter vector. Poor excitation leads to a large variance of the impulse response estimate, resulting in a slowly converging adaptive algorithm. The standard solution is to add a scaled identity matrix to the ill-conditioned input correlation matrix, where scaling is performed with an estimate of the near-end background noise power. We illustrate how this type of regularization fits in a linear minimum mean square error framework and how regularization may be improved by incorporating prior knowledge on the room impulse response. Prior knowledge can be constructed based on some physical parameters of the acoustic setup. Offline simulation results indicate that the proposed regularization technique may yield a low-variance room impulse response estimate.