Individuation of postlexical phonology for speech synthesis

Postlexical phonology exhibits a great deal of interspeaker variation, even within widely construed dialects such as American English. Contemporary speech synthesizers have concentrated on modeling the acoustic properties of individual speakers, but not necessarily their phonologies. We discuss a method for inferring individual postlexical phonologies from labeled corpora. Using this method, individualized phonologies can be combined with individualized acoustic models, thereby enabling a speech synthesizer to achieve a closer facsimile of an original voice.