Adjacent node continuous-state HMM's

This paper explores properties of a family of Continuous state hidden Markov models (CSHMM’s) that are proposed for use in acoustic modeling. These models can be viewed as applying a smoothing to ordinary HMM’s in order to make estimates of transition and observation probabilities more robust by sharing data between adjacent state nodes. They may be trained by EM so that all parameters properly reflect the applied smoothing. The amount of smoothing may be trained as well, and the model reverts to the ordinary HMM in the limit as the smoothing parameters reach zero. Thus this technique may be employed selectively only in areas in the model where training data is sparse. This paper formulates EM-training for one variant of these models, and explores their performance when applied to constructing ergodic CSHMM models of speech and to a phoneme recognition task on the same data. The ergodic CSHMM did not improve performance over the HMM, but the phoneme CSHMM’s model the data with higher likelihood than the equivalent HMM’s, and have superior recognition accuracy.