Abstract:

We present two methods for building a finite-state transducer which generalizes a finite-state transduction to any word on a given alphabet. The methods are exact in the sense that the inferred transducer coincides on the inputs for which the initial function is defined. We apply the methods to the problem of grapheme-to-phoneme transcription. In this case, the initial function is given in the form of an aligned letters--phonemes dictionary. The generalized function gives a phonetic transcription for words not in the dictionary based on its similarity to other words. In addition, for this problem, the generalization could be represented more compactly than the initial dictionary.