Generalization of state-observation-dependency in Partly Hidden Markov Models

Generalized Partly Hidden Markov Model (GPHMM) is proposed by modifying Partly Hidden Markov Model (PHMM), and it is successfully applied to the speech recognition. PHMM, which was proposed in our previous paper, is the novel stochastic model, in which the pairs of the hidden states (H-state) and the observable states (O-state) determine the stochastic phenomena of the current observation and the next state transition. In the formulation of PHMM, we used common pair of H-state and O-state to determine both of these phenomena. In the formulation of GPHMM proposed here, we use common H-state but different O-states for the current observation and for the next state transition separately. This slight modification brought the big flexibility in the modeling of phenomena. Experimental results showed the effectiveness of GPHMM (without delta parameters): it reduced the word error by 17% compared to triphone HMM (with delta parameters), respectively.