摘要 |
PURPOSE: A method for decreasing memory in successive sound recognizer is provided to efficiently decrease a memory by setting three constraint conditions and tying states without lowering of recognition rate. CONSTITUTION: Each cluster to each state of a HMM(Hidden Markov Model is set(S10). A pair of clusters are set(S20). The number of occurrence frequency is checked during training the number of occurrence frequency(S30). A clustering is not performed when the number of occurrence frequency is over a threshold value as a result of the check, and then the step(S20) is repeatedly performed. Three restraint conditions are applied when the number of occurrence frequency is less the threshold value as a result of the check and the clustering is preformed(S40). Three restraint conditions have a first condition for tying a same allophone among center states, a second condition for tying an allophone occurred in a same phoneme, and a third condition for tying states in a same position. |