Text Normalization for the Pronunciation of Non-standard Words in an Inflected Language

In this paper we present a novel approach, called “Text to Pronunciation (TtP)”, for the proper normalization of Non-Standard Words (NSWs) in unrestricted texts. The methodology deals with inflection issues for the consistency of the NSWs with the syntactic structure of the utterances they belong to. Moreover, for the achievement of an augmented auditory representation of NSWs in Text-to-Speech (TtS) systems, we introduce the coupling of the standard normalizer with: i) a language generator that compiles pronunciation formats and ii) VoiceXML attributes for the guidance of the underlying TtS to imitate the human speaking style in the case of numbers. For the evaluation of the above model in the Greek language we have used a 158K word corpus with 4499 numerical expressions. We achieved an internal error rate of 7,67% however, only 1,02% were perceivable errors due to the nature of the language.

• Type: Conference
• Year: 2004
• Where: SETN
• Authors: Gerasimos Xydas, Georgios Karberis, Georgios Kouroupetroglou
• Keywords: Artificial Intelligence, Augmented Auditory Representation, Human Speaking Style, Proper Normalization, SETN 2004