| 摘要 |
PURPOSE:To improve recognition accuracy at high speed by performing recognition processing on a recognition target satisfying a condition with respect to the continuous length of a unit template. CONSTITUTION:The feature quantity analyzed by a voice analysis part 102 of voice converted to an electrical signal is extracted. Meanwhile, the text candidate of the recognition target is generated at a text generating part 106 by using the connection rule of the word of a word storage part 103 and a grammar storage part 104, and the minimum voice emission time and the maximum voice emission time of a continuous length condition storage part 105. At a scoring part 107, the recognition target of the word is decided by input feature quantity and the maximum continuous length, and scoring is applied to the feature quantity of the text candidate, and a recognition result is outputted from a selection part 108. In such a way, the recognition accuracy and speed can be improved by setting each unit template only when it is provided with the most likelihood. |
