Grapheme-to-phoneme Conversion in Theory and Practice

Tools for mapping between written words and phonetic forms are essential components in many applications of speech technology, such as automatic speech recognition (ASR) and speech synthesis (TTS). Simple converters can be derived from annotated speech corpora using machine learning, and such tools are available for almost all European languages and a great number of others. Whereas their performance is adequate for ASR and for low-quality TTS, their lack of precision makes them unfit for linguistic research purposes such as phonetic annotation of spontaneous speech recordings. A common method of enhancing their predictive power (e.g. faced with out-of-vocabulary tokens) is to include phonetic and lexical rules, and sometimes even semantic and contextual knowledge. In this paper we present some of the principles underlying the typical linguistically informed phonetic converter. We illustrate our points with examples from the Danish grapheme-to-phoneme converter Phonix.