A unified view for discriminative objective functions based on negative exponential of difference measure between strings

摘要: This paper presents a novel unified view of wide variety objective functions suitable for discriminative training applied to sequential pattern recognition problems, such as automatic speech recognition. Focusing on central component conventional functions, the sum modified joint probabilities observations and strings, analysis generalizes these by weighting each term in an important function, negative exponential difference measure between strings. The interesting valuable results this investigation are highlighted comprehensive relationship chart that covers all common approaches (Maximum Mutual Information, Minimum Classification Error, Phone/Word Error), well corresponding generalizations modifications those approaches.