Automatic generation of phonetic regression class trees for MLLR adaptation

In this paper, it is shown that a correlation criterion is the appropriate criterion for bottom-up clustering to obtain broad phonetic class regressio n trees for maximum likelihood linear regression (MLLR)-based speaker adapt ation. The correlation structure among speech units is estimated on the spe aker-independent training data. In adaptation experiments the tree outperfo rmed a regression tree obtained from clustering according to closeness in a coustic space and achieved results comparable with those of a manually desi gned broad phonetic class tree.