 |
Modeling Pronunciation Variation for Automatic Speech RecognitionRolduc, The Netherlands |
  |
|
Modeling Pronunciation Variation for Automatic Speech RecognitionRolduc, The Netherlands |
|
Pronunciation variations are common sources of recognition errors in real-world applications, so that specific techniques must be developed to handle them. We are describing a method to incorporate pronunciation alternatives that have been tested with both continuous and isolated word speech recognisers for Spanish. We present an automatic grapheme-to-phoneme system, modified to generate alternate pronunciations. It works according to phonological rules manually developed using certain variations, well known in the linguistic community but not widely exploited in the Spanish speech recognition arena. We will apply this strategy only to the recognition stage of both a continuous speech recogniser for clean speech data, and an isolated one for a telephone environment task. We will report improvements up to 20% decrease in error rate, for the continuous speech task, while for the isolated word recognition task, no significant effect has been found. We will conclude analysing which effects have led to these results and discuss future work to be done.
Bibliographic reference. Ferreiros, Javier / Macias-Guarasa, Javier / Pardo, José M. / Villarrubia, Luis (1998): "Introducing multiple pronunciations in Spanish speech recognition systems", In MPV-1998, 29-34.