摘要 |
PURPOSE: To easily form the neural net at a high recognition factor without collecting many talkers for every car which is different in acoustic characteristics or collecting the voice data. CONSTITUTION: The standard voice is previously recorded by a recorder 4 in anechoic room 1. This standard voice is played back by a player 7 in the cage to be really used for voice recognition. This voice is inputted to a preprocessing unit 11 through a microphone 9 and an A/D converter 10 so as to form the input pattern 12. This input pattern 12 is inputted to a neural net 13, and the output pattern 14 is outputted as a result of recognition. On the other hand, the teaching signal 16 is given from outside so as to be compared with the output pattern 14, and on the basis of the error 15 between them, each combination weight of the neural net 13 is corrected. This operation is repeated of all standard voices already recorded. |