ISCA Archive Eurospeech 1993
ISCA Archive Eurospeech 1993

Hidden Markov models using shared vector linear predictors

B. A. Maxwell, Phil C. Woodland

It has been previously shown that augmenting a standard HMM with a set of vector linear predictors can improve recognition rates compared with standard HMMs. The set of vector linear predictors associated with each state improve the HMMs ability to model the correlations in real speech data, and help to overcome the HMM state-conditional independence assumption. However, introducing extra parameters into the model requires more training data. This problem can be partly overcome by sharing the predictor parameters between multiple HMM states, and hence more robust, but less specific estimates of the predictor parameters are obtained. This paper develops the theory and im- plementation of arbitrarily shared vector linear prediction for hidden Markov models. For most predictor offsets, predictors shared across all states of an HMM provide more accurate recognition on both training and test data sets than equivalent HMMs without predictors when evaluated on a British English E-set recognition task.

Keywords: Speech recognition, vector prediction