| 摘要 |
PURPOSE: A method of dividing the initial state of voice data using hidden markov model is provided to split the state of voice data using IS cumulative distance between neighboring frames so as to converge an initial state transfer probability parameter to the global maximum value. CONSTITUTION: A pre-processing is carried out(S1,S2,S3). IS cumulative distance is divided by the number of states(S4). The initial state of the divided value is split(S5). An initial state transfer probability parameter is determined using the split states(S6). The initial state transfer probability parameter is reestimated(S7). The state is split again by performing viterbi algorithm(S8). A model parameter is stored when a model converges(S10) and the parameter is reestimated until the model converges when the model does not converge.
|
