Room Impulse Response Estimation with Kernel-Based Regularization

Abstract

This paper discusses the identification of the room impulse response, which is an acoustic model of the sound reflection along the room. In particular, we consider the case where the input sound lacks some frequency elements. Such a situation occurs due to the transfer characteristics of the speaker, and it leads to overfitting behavior. To avoid this overfitting, we use the kernel-based regularization which enjoys a priori knowledge on the target system. Room impulse responses are much longer than the ones considered in the existing kernel-based regularization methods, thus we use orthonormal basis function expansion to reduce computational burden. We show a numerical study with the impulse response measured in a real concert hall, and confirm that the kernel-based regularization improves the identification accuracy than the least square method.