摘要 |
PURPOSE: To provide a learning method and a device therefor minimizing the expected value of a correct answer ranking in a ranked recognition result outputted from a signal recognition device. CONSTITUTION: This method is a learning method for a recognition parameterΛ1 of a signal recognition device, and by inputs a known signal X, obtains a first class of score gX and one or plural other classes of scores g1 -gW (except gX) using the known signal X and the recognition parameterΛ1 obtains determining ranking measure signals d1 -dW from these scores g1 -gW, and by using the known signal X and the ranking measure signals d1 -dW, in a HMM(hidden Markor model) updating device 88, corrects a recognition parameter of a desired class so as to reduce an expected value of a ranking number of a correct answer result in a recognition result of an unknown signal.
|