Canonical correlation based impersonation quality determination algorithm for natural morphed speech

In this article, impersonation experiments were conducted utilizing natural morphed speeches between (/a/-/b/-/a/) and (/a/-/g/-/a/) as stimuli. The sound stimuli are produced utilizing the natural glottal source and morphed linear predictive coding (LPC) filtering coefficients, which represent the vocal tract states. An algorithm has been proposed for determination of impersonation quality based on the redundancies of the canonical variates for the morphed sounds. The redundancy is defined by canonical variate loading factor and corresponding canonical correlation coefficient. The quality of imitated speech for different impersonators is evaluated by applying the proposed algorithm.