Functional Pronunciation Dictionaries

This document describes the system Functional Pronunciation Dictionaries (FPD), a language-independent system for defining pronunciation lexicons. The starting point of this system is Functional Morphology (FM) [3, 2], where we asked us the question — given that we already have defined a lexical resource in FM, what would be an efficient approach to extending the lexicon with high-quality pronunciation information? We will use Swedish to illustrate the ideas of FPD. The main idea of FPD is based on the assumption that we can get high correctness by using automatic transcription, and if a transcription is erroneous, it is normally not completely wrong — it may be one or two sounds that are incorrect. Given that this assumption is true, then automatic transcription is a reasonable approach as the first step in the creation of a high-quality pronunciation dictionary. The lexicographer’s job would be to adjust the incorrect transcriptions. M. Uneson [8] reports a precision of 95.7% in his automatic transcription system of Swedish, i.e. approximately every twentieth word is transcribed wrongly. Uneson gives a list of the most problematic cases: compounds; lexical stress; lexical word accent; loan words; pronunciation that is not derivable by orthography; occasional exceptions to normal orthographical markings. Moreover, he reports that the main problem is compound resolution (the compound boundaries are unmarked in Swedish). The compound resolution is, of course, connected to the problem of assigning the lexical stress.