Multilingual Data-Driven Pronunciation

Automatic pronunciation of unknown words is a hard problem of great importance in speech technology. The difficulty of the problem appears to vary across languages, according to the ‘depth’ of orthography, although there are, as yet, little by way of quantitative comparisons. Early solutions were based on manually-written expert rules but these are expensive to derive and maintain, are entirely language-specific and—according to more recent evaluations using large test sets—do not perform very well (at least for English). By contrast, data-driven methods (which infer pronunciations from a set of examples) require little expert knowledge, can be quickly and easily applied to new languages for which an appropriate set of examples is available, and seem to perform far better than rules. In this paper, we compare the success of data-driven automatic pronunciation for four European languages: English, French, Frisian and German. We also attempt to quantify the difficulty of the problem using the entropy of the association between letters and phonemes found after prior alignment of the various dictionaries.