Speech recognition system, training system and method for the calculation of iteration values for free parameters of a maximum entropy speech model - -

摘要

The present invention relates to a voice recognition system and a method for the calculation of iteration values for free parameters λ alpha of the maximum entropy speech model - -. It is known from the prior art, this free parameters λ alpha, for example, with the aid of a cyclically iteratively - gis training algorithm to approximate. Cyclically in this case means that in each iteration step, n, a cyclically predetermined attribute group ai (n) of the speech model for calculation of the n + 1 iteration value for the free parameters is evaluated. Such a rigidly cyclically associated attribute group ai (n) is, however, not always the most suitable, the gis training algorithm - in a current situation at the fastest and most efficient to converge. It is therefore, according to the invention, a process for the selection of the in this respect, the most suitable attribute group is proposed, wherein the degree of adaptation of the iteration edge values dollar i1 to in each case associated desired boundary values m alpha for all the features of the respective attribute group, as a criterion for the selection of the attribute group serves.