Estimation of acoustic resonances for room transfer function equalization

Strong acoustic resonances create long room impulse responses (RIRs) which may harm the speech transmission in an acoustic space and hence reduce speech intelligibility. Equalization is performed by cancelling the main acoustic resonances common to multiple room transfer functions (RTFs), i.e., common-acoustical-poles, in the room. This paper discusses the utilization of different norms (i.e., 2-norm and 1-norm) and models (i.e., all-pole and pole-zero) for RTF modelling and then equalization. Acoustic resonances may be modelled by means of the poles of the RTF. In the literature, however, it is not clear what model (i.e., all-pole or pole-zero) generates pole estimates that perform better in RTF equalization. Furthermore, least squares error (i.e., 2-norm) minimization is typically employed for the estimation of the poles. In this paper a least absolute error (i.e., 1-norm) minimization is further proposed for pole estimation. A comparative evaluation of these different norms and models in terms of their residual RTF and residual RIR (i.e., the residuals after equalization) is provided.