Handling dictation ambiguities in the production of text from large-vocabulary speech recognition