Representation of large lexica using finite-state transducers for the multilingual text-to-speech synthesis systems