We adopted the Korean phonological rules to state clustering ofcontextual domain for representing the unknown contexts and tying themodel parameters of new states in state clustering of SSS (successivestate splitting). We used the decision tree-based successive statesplitting (DT-SSS) algorithm, which splits the state of contexts basedon phonetic knowledge. The SSS algorithm proposed by Sagayama (1992) isa powerful technique, which designed topologies of tied-state HMMsautomatically, but it does not generate unknown contexts adequately. Inaddition it has some problem in the contextual splits procedure. In thispaper, speaker independent Korean isolated word and sentence recognitionexperiments are carried out. In word recognition experiments, thismethod shows an average of 6.3% higher word recognition accuracy thanthe conventional HMMs, and in sentence recognition experiments, it showsan average of 90.9% recognition accuracy
展开▼
