TABLE OF CONTENTS Text pre-processing involves turning raw texts (possibly composed of abbreviations, acronyms, numbers, dates and telephone numbers) into a normalized written form, acceptable by the synthesis system. Once a particular structure has been detected, two approaches are possible:
replace the raw character string by a normalized letter string which corresponds to the pronunciation of the acronym or the number. This is the approach adopted by most text pre-processing systems; or
carry the detected structure as far as possible through the sequence of linguistic processing. If, for example, a sequence of numbers is analyzed as a telephone number, specific intonation can be applied by the prosody generation module.
Pre-processing operates at the interface between a general-purpose synthesis system and an application that is necessarily targeted to a particular service. Thus, pre-processing names and proper nouns in a reverse directory service [5] is different to pre-processing a fax or e-mail for a reading application.
The most common technical solution to this problem involves scanning exception lexicons containing (for each acronym or abbreviation) a normalized word string which corresponds to a non-ambiguous pronunciation [6]. Detection of structure such as currencies, telephone numbers or dates is generally achieved by rules. It must however be noted that some pre-processing systems rely on probabilistic approaches such as n-grams.
Grapheme-to-phoneme conversion or transcription involves determining the sequence of phonemes corresponding to the pronunciation of a written text.
The main difficulty in grapheme-to-phoneme...
