| 摘要 |
PROBLEM TO BE SOLVED: To improve the recognition rate by going back to the past and redoing a recognition calculation when it is considered that a correct answer hypothesis is not contained in a beam in a frame synchronous beam search. SOLUTION: In a continuous voice recognition, this voice recognition device is provided with a hypothetical memory judging means 1 to judge whether to store a hypothesis or not, a hypothesis storage means 2 to store a past hypothesis, a search control means 4 to control a frame synchronous beam search, a search calculating means 5 to execute calculation of the frame synchronous beam search, and a correct answer judging means 3 to investigate a possibility that a correct answer hypothesis is contained in the beam. And, when the correct answer hypothesis is judged to be contained in the beam, the speech recognition device stores the hypothesis, and when the correct answer hypothesis is not judged to be contained in the beam, the device uses a hypothesis stored in the past to go back to a past frame, and expands the beam width for redoing the search. |
