| 摘要 |
PURPOSE:To improve the quality of a synthetic speech without using any parametric model by determining time length assigned to a phoneme by a neural network. CONSTITUTION:Time distribution to units constituting a speech is determined by the neural network. Namely, an output 104 of each element xij 103 of an input layer 102 is coupled with each element 106 of a 2nd layer 105 and an output 107 of each element 106 of a 2nd layer 105 is coupled with each element 109 of a 3rd layer 108. An output 110 of each element 109 of the 3rd layer 108 is coupled with an element 112 of an output layer 111. Then when factors of respective phonemes are inputted as an input 101, time length T1 which is distributed to input phonemes is outputted 113 successively as the output of the element 112 of the output layer. Consequently, a synthetic speech of excellent speech quality which has no mechanical unnaturalness is obtained. |
