LINEAR PREDICTION BASED DIFFUSE SIGNAL ESTIMATION FOR BLIND MICROPHONE GEOMETRY CALIBRATION

Spatial cross coherence function between two locations in a diffuse sound field is a function of the distance between them. Earlier approaches to microphone geometry calibration utilizing this property assume the presence of an ambient noise source. Instead, we consider the geometry estimation using a single acoustic source (not noise) and show that late reverberation (diffuse signal) estimation using multi-channel linear prediction (MCLP) provides a computationally efficient solution to geometry estimation. The idea behind this is that, the component of a reverberant signal corresponding to late reflections satisfies the diffuse sound field properties, which we exploit for distance estimation between microphone pairs. MCLP of short-time Fourier transform (STFT) coefficients is used to decompose each microphone signal into early and late reflection components. Cross coherence computed between the separated late reflection components is then used for pair-wise microphone distance estimation. Multidimensional scaling (MDS) is then used to estimate the geometry of the microphones from pair-wise distance measurements. We show that, higher reverberation, though detrimental to signal estimation, can aid in microphone geometry estimation. Estimated position error of less than 2 cm is achieved using the proposed approach for real microphone recorded signals