Speaker adaptation by a linear transformation with optimised parameters

Speaker dependence of automatic speech recognition systems can be reduced by applying speaker-specific transformations to adapt the speech signal of a new speaker to that of the reference speaker. Initial investigations showed that speaker adaptation can be performed by transformations using spectral weighting and spectral warping. These heuristic methods can be substituted by a general linear matrix transformation, the parameters of which are determined by mean square error optimisation. The improvement of the recognition rate achievable by this matrix transformation is very high, but the method needs a large learning set. This can be reduced by restriction of the matrix to a band including the main diagonal in the middle. This banded matrix yields results close to those of the general matrix. Adaptation can be performed speaker-specifically as well as speaker- and class-specifically. As the cost of phoneme class-specific adaptation is very high, a grouping of phonemes is proposed so that one adaptation parameter set is used for all phonemes that belong to any one group.