Joint uncertainty decoding with unscented transform for noise robust subspace Gaussian mixture models

Common noise compensation techniques use vector Tay-lor series (VTS) to approximate the mismatch function. Recent work shows that the approximation accuracy may be improved by sampling. One such sampling technique is the unscented transform (UT), which draws samples deterministically from clean speech and noise model to derive the noise corrupted speech parameters. This pa-per applies UT to noise compensation of the subspace Gaussian mixture model (SGMM). Since UT requires rel-atively smaller number of samples for accurate estima-tion, it has significantly lower computational cost com-pared to other random sampling techniques. However, the number of surface Gaussians in an SGMM is typi-cally very large, making the direct application of UT, for compensating individual Gaussian components, compu-tationally impractical. In this paper, we avoid the com-putational burden by employing UT in the framework of joint uncertainty decoding (JUD), which groups all the Gaussian components into small number of classes, shar-ing the compensation parameters by class. We evaluate the JUD-UT technique for an SGMM system using the Aurora 4 corpus. Experimental results indicate that UT can lead to increased accuracy compared to VTS approx-imation if the JUD phase factor is untuned, and to similar accuracy if the phase factor is tuned empirically. 1