| 摘要 |
PURPOSE: To provide the method for determining the word sequence of the specified vocabulary of a speech signal which conducts a little danger of the loss of a correct sequence of words during comparison while keeping search effort still relatively a little. CONSTITUTION: The speech signal is recognized until the output of a recognized word sequence or recognized sentence is divided. Only word hypotheses are generated individually at different start moments, and preliminary word strings are formed from the individual word hypotheses according to a word graph, which is continuously optimized by erasing the part of the word string. For the purpose, parts of word strings having the same start point and end point are compared with each other and the score of a word having a cycling end point is compared with a threshold value. Postediting operation is introduced as the optimization step of the word graph. In this case, other words having the same start are all outputted for wrong words and an operator speedily selects a correct word out of all other words. |
