Data-driven phonetic regression class tree estimation for MLLR adaptation

In this paper a method is presented to estimate a broad phonetic class regression tree to be used in MLLR adaptation. The tree is derived from the correlation structure among phone units estimated on the training data. The algorithm is language-independent and showed good results on both an English and a Mandarin Chinese database. In adaptation experiments the tree outperformed a regression tree obtained from clustering according to closeness in acoustic space and achieved results comparable with those of a manually designed broad phonetic class tree.