Structural maximum a posteriori linear regression for fast HMM adaptation

Transformation-based model adaptation techniques like maximum likelihood linear regression (MLLR) rely on an accurate selection of the number of transformations for a given amount of adaptation data. If too many transformations are used, the transformation parameters may be poorly estimated, can overfit the adaptation data, and offer poor generalization. On the other hand, if the number of transformations is too small, the adapted models can only provide a moderate improvement over the baseline models. An adaptation approach should therefore be flexible in order to estimate reliably a large number of transformations when the amount of adaptation data is large, and a small number of transformations when only a few adaptation utterances are available. In this work, we show that a significant improvement can be obtained over MLLR with dynamic regression classes, first by replacing the maximum likelihood estimation criterion by a maximum a posteriori criterion, then by introducing a tree-structure for the prior densities of the transformations. The effectiveness of the proposed approach is illustrated on the Spoke3 1993 test set of the WSJ task. Using the same regression classes as MLLR, it is shown that the proposed approach reduces the risk of overfitting and exploit the adaptation data much more efficiently than MLLR, leading to a significant reduction of the word error rate with as little as one adaptation utterance.