Sixth European Conference on Speech Communication and Technology
We propose a novel hierarchical mixture model and present its application to acoustic modeling for HMM based large vocabulary conversational speech recognition. We detail an EM algorithm for estimating the parameters of such a mixture tree for the case of Gaussian component densities. We sketch how clustering algorithms can be applied to automatically construct suitable mixture trees for a large number of HMM states. Furthermore, we discuss the advantages of a tree structured organization of mixture densities in terms of effciency of evaluation and scalability of context modeling. These properties allow (1) to arbitrarily downsize trained mixture trees thereby trading o recognition accuracy against decoding speed and model size and (2) to adapt the structure of trained mixture trees in cross domain applications to re ect the di ering requirements in speci city of context. We present preliminary results of using mixture trees for recognition experiments on the Switchboard LVCSR corpus.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Fritsch, J. (1999): "Mixture trees - hierarchically tied mixture densities for modeling HMM emission probabilities", In EUROSPEECH'99, 1103-1106.