| 摘要 |
<p>Speech input to input terminal (110) is converted in analyzer (120) into a feature vector series, which is then fed to input pattern memory (120) to be held as an input pattern and is also fed to a preliminary recognizor (160). The preliminary recognizor (160) executes preliminary recognition by using the input pattern and all reference patterns stored in reference pattern memory (150), thus obtaining top N candidates in the order of higher similarities. A reference pattern adapter (170) executes the adaptation of reference patterns by using the input pattern, the top N candidates as a result of the preliminary recognition and corresponding reference patterns, the result being stored in reference pattern memory (150). A final or second recognizor (180) executes re-recognition of the input pattern by using the adapted reference patterns corresponding to the top N candidates obtained as a result of the preliminary recognition, the result being output to an output terminal (190). <MATH></p> |
