IMPROVEMENT OF MAP-VFS ADAPTATION PERFORMANCE BY FUZZY CONTROL

摘要

This work presents a fuzzy control mechanism for the conventionally adopted maximum a posteriori vector field smoothing (MAP-VFS) speaker adaptation scheme. The proposed mechanism, called FC-MAPVFS, regulates the influence of MAP-VFS adaptation when the training data from a new speaker is inadequate. FC-MAPVFS flexibly manages both the calculation of the weight control parameter from MAP and the estimation of interpolated transfer vectors from VFS based on the amount of adaptation data, thus ensuring that the MAP-VFS adaptation is robust against data scarcity. The proposed mechanism is conceptually simple and effective. Experimental results indicate that FC-MAPVFS outperforms conventional MAP-VFS, particularly when the adaptation data are scarce.