SPEAKER LOCALIZATION USING FROBENIUS NORM WITH A FOCUS ON CLOSE SPEAKER AND NOISE SOURCE

Speaker direction of arrival (DOA) estimation is an important task for beamforming-based noise reduction and camera steering. Common DOA estimation techniques suffer from biased DOA estimates when the speaker is close to the noise source. In this paper, a novel speaker DOA estimation is presented based on Frobenius norm minimization. In our model, multiple possible speakers are located in each one of a predefined set of candidate DOAs. Instead of estimating the DOA, the power spectral density (PSD)s of the speakers are mutually estimated and the dominant DOA is determined by the speaker with the maximal PSD. The PSDs estimation task is then employed by minimizing the Frobenius norm of the matrix-difference between the estimated PSD matrix of the received signals and the model-matrix described the multiple speaker presence. An experimental study demonstrates the benefits of the proposed Frobenius-based DOA algorithm in simulated dataset w.r.t. a maximum likelihood (ML) based DOA estimator, especially when the speaker is angularly close to the noise source.