DanPO-a transcription-based dictionary for Danish speech technology

We present a new strategy for the creation of phonetic lexicons. As we argue, lexical resources for speech technology integration should be informed by transcriptions of spontaneous speech. We illustrate our strategy with examples from the dictionary DanPO (Danish Phonetic-Orthographic Dictionary) which is developed at the Center for Computational Modelling of Language (CMOL). For reference corpus we used DanPASS consisting of 57 recordings of task-oriented monologs, transcribed by professional and MAlevel phoneticians using the Danish SAMPA phonetic alphabet. From the transcriptions, dictionaries and concordances were compiled, and these resources were merged with the (prescriptive) phonetic renderings of a standard Danish word dictionary of 87,000 lemmata. As an effect of the “transcription informed” strategy, DanPO is expected to significantly improve the success rate of automatic speech recognizers, as well as the naturalness of artificial voices. Furthermore, we devise an experimental strategy in order to evaluate the dictionary and further improve later versions.