High precision robust modeling of long room responses using wavelet transform

Abstract

Modeling of a room impulse response (RIR) is required in many audio processing applications; however, this is challenging since room responses are usually long and complex in practice and drastically vary as the source and microphone locations change. In this paper, a subband multichannel modeling method is proposed, which is computationally efficient, precise, and robust against RIR variations. A dual-tree complex wavelet packet transform is utilized to decompose a multichannel RIR into aliasing-free subband signals, and low order adaptive Kautz filters are designed to model subband signals using the poles common to the RIR channels. A least-squares algorithm is introduced to efficiently estimate the common poles at each subband. Experimental results indicate that the proposed method accurately models long room responses, while exhibiting significant robustness against room response variations caused by changing the source and microphone locations.