The speech recognition system by the largest entropy language model

摘要

As for this invention, with help of training algorithm before it regards the method of setting this parameter when it is not possible to decide the free orthogonalization parameter of the attribute in the largest entropy audio model. One purpose of this invention, speech recognition system 100, training device, is to offer the method of deciding the kind of parameter ^{ortho which} possesses 10, and many possible interpretations. This purpose is achieved by calculating on the basis of this invention, formula (i) formula (ii) and denominator (denominator) (= formula (iii)) and also.