Analogy, the Corpus and Pronunciation

Abstract

Reading aloud and text-to-speech synthesis share the commonality of taking a printed text and generating the acoustic correlate. This chapter starts from the premise that modelling the human solution for generating the pronunciation of a word previously unseen will result in more effective and accurate pronunciation modules for machine synthesis. The human and the text-to-speech system both have access to a lexicon: the mental lexicon in the first case and the system dictionary in the other. These are the data sources which can be mined to generate an unknown word’s pronunciation. This chapter presents a computational approach to automatic pronunciation developed from a psychological model of oral reading. The approach takes the system dictionary — a frequency-tagged corpus — and uses analogy to generate the pronunciation of words not in the dictionary. A range of implementational choices is discussed and the effectiveness of the model for (British) English, German and Māori demonstrated.