Preference-Driven Bimachine Compilation : An Application to TTS Text Normalisation
Wojciech Skut · 2004
This paper describes a grammar formalism and a deterministic parser developed for text normalisation in the rVoice1 text-to-speech (TTS) system. The rules are formulated using regular expressions and converted into a non-deterministic finite-state transducer (FST). At runtime, search is guided by parsing preferences which the user may associate with regular operators; the best solution is determined in a way similar to the directional evaluation of constraints in Optimality Theory. During compilation, the FST is converted into a bimachine, making deterministic parsing possible.