Speech segmentation using a hypothesis test based on Random Matrix Theory

Speech segmentation to covariance-stationary regions is of interest, for example in subspace-based speech enhancement. However as the true covariance matrices of speech segments are unknown, it is usual to use their sample estimates. To check whether two sample covariance matrices have been drawn from the same distribution or not, we have used a test statistic previously proposed for image segmentation. We have derived a new expression for the decision threshold using Random Matrix Theory. Finally, a novel segmentation procedure is proposed and applied to both synthetic and speech data. The presented simulation results show the low computational cost and good performance.