7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Maximum Mutual Information Training of Hidden Markov Models with Vector Linear Predictors

K. K. Chin, P. C. Woodland

Cambridge University, U.K.

HMM makes a piece-wise constant assumption about the temporal evolution of the speech signal. This is not true for speech signals which are known to be highly temporally correlated. Many approaches had been proposed to overcome the limitation of HMMs in modelling temporal context. One of these approaches uses a Vector Linear Predictor (VLP) to model the relationship between a nearby frame and the current frame. In this paper, Maximum Mutual Information Estimation (MMIE) of VLP-HMMs is explored. The MMIE training of VLP-HMMs are evaluated on WSJ data.


Full Paper

Bibliographic reference.  Chin, K. K. / Woodland, P. C. (2002): "Maximum mutual information training of hidden Markov models with vector linear predictors", In ICSLP-2002, 997-1000.