Bayesian regularization and nonnegative deconvolution (BRAND) for acoustic echo cancellation

摘要

The Bayesian regularization and nonnegative deconvolution (BRAND) algorithm is used to estimate the acoustic room impulse response by incorporating prior information such as sparsity and nonnegativity about the filter coefficients. For experimental measurements with microphones and speakers with non-ideal characteristics, the overall transfer function can be decomposed into a common short FIR filter, and a long nonnegative filter representing the room response. We develop an online estimation procedure for the BRAND algorithm, along with a computationally efficient implementation. Simulations and experimental results show the robustness of the resulting algorithm for echo cancellation in the presence of large ambient noise.