Speaker normalization based on time-frequency warp with inter-frame consistency

A new algorithm for speaker-independent spoken word recognition is presented. The algorithm is based on the time-frequency warping technique where frequency axis warping is performed in order to adjust individual spectral difference, in addition to time axis warping. In the conventional algorithm, frequency axis warping is independently determined at each frame (i.e., time). In this case, such warp have a tendency to yield excessive deformations of time-frequency plane, it is feared. In order to suppress such excessive deformations, inter-frame consistency of frequency axis warping is newly taken into account as constraints on the warping. The optimal warping is obtained by using dynamic programming with the constraints. As an implementation technique, beam search based acceleration is also investigated. Experimental results indicates advantageous characteristics of the present algorithm over the conventional algorithm.