| 摘要 |
PURPOSE:To execute correct recognition with a small processing quantity even when variance between patterns is large and similar patterns are many by not using always a KNN method for the recognition and using the KNN method in the appropriate case. CONSTITUTION:In a registration mode, a sound signal inputted from a microphone 1 is sound-analyzed by a sound analyzing part 2, the time series pattern of respective standard patterns is stored into a standard pattern memory part 4, and by using a KNN method, the number of the pattern to be recognized is stored into a KNN pattern table 6. In a recognition mode, a matching part 8 calculates a DP distance for two time series patterns from an input pattern memory 7 and the memory part 4 by using a DP matching method and sends it to a recognizing part 6. The recognizing part 9 sends the number of the standard pattern, in which the DP distance comes to be minimum, and the category name to a deciding parrt 10. The deciding part 10 collates the number with a table 6, when the KNN method is not used, outputs the category name as the recognizing result, and when the method is used, sends the category name and it in the table 6 to a KNN processing part 11.
|
