On transcribing informally-pronounced numbers in Romanian speech

The pronunciation model, a mapping between the lexicon words and their phonetic representation, has a key role in automatic speech recognition. Although many times neglected, the accuracy of this model influences significantly the accuracy of the whole system. This study discusses within-word and cross-word pronunciation variations for Romanian numbers and proposes the solutions to model them in the phonetic dictionary and the language model of an existing speech recognition system for Romanian. The evaluation is performed of a read speech corpus comprising rational numbers with up to three decimal digits. The experiments show a relative WER improvement of 14% over the baseline when within-word pronunciation variations are taken into account and an additional relative WER improvement of 63% when cross-word pronunciation variations are also modelled.