Context modeling with the stochastic segment model

The authors describe an approach, the stochastic segment model, for context modeling in continuous speech recognition for models based on multivariate Gaussian distributions. Typically, robust context models in hidden Markov models (HMMs) are obtained by using mixture distributions; here the authors tie covariance parameters across classes of similar context. The specific classes over which parameters are tied can be based on models with less context or determined by clustering, where they have investigated both hand-specified linguistically motivated clusters and automatic k-means clustering. Experimental results on phoneme classification show that clustering improves performance, and word recognition results show that error reduction over context-independent models using this approach is comparable to that achieved with discrete hidden-Markov models using mixture distributions