 |
Modeling Pronunciation Variation for Automatic Speech RecognitionRolduc, The Netherlands |
  |
|
Modeling Pronunciation Variation for Automatic Speech RecognitionRolduc, The Netherlands |
|
The role of a machine-readable pronunciation dictionary in ASR typically involves the look-up at some stage of a canonical pronunciation represented in phonemic transcription. Beyond this basic information, the problem of intra- and inter-speaker variation in usually thought to require the application of general rules which use phonetic knowledge about pronunciation variation to supply appropriate variant forms for different phonological contexts and different speakers. This approach is supported by the fact that most machine-readable dictionaries provide only one pronunciation per word, except where there are widely differing alternatives, and therefore contain little or no information about variation. However, specialist pronunciation dictionaries exist for English which contain in electronic form a rich documentation of pronunciation variants in the form of a computer database, and this in principle makes the information available for use in ASR systems. Unfortunately, the coding of information about alternative pronunciations for human users of the dictionaries is not easily adaptable for ASR purposes, and the number of alternatives coded in these dictionaries has never been calculated. This paper discusses how the information compactly coded in a pronouncing dictionary might be expanded into a full list of regularly found pronunciation variants which retains the details of preferences and contextual forms and could be used in ASR applications.
Bibliographic reference. Roach, Peter / Arnfield, Simon (1998): "Variation information in pronunciation dictionaries", In MPV-1998, 121-124.