 |
Modeling Pronunciation Variation for Automatic Speech RecognitionRolduc, The Netherlands |
  |
|
Modeling Pronunciation Variation for Automatic Speech RecognitionRolduc, The Netherlands |
|
In this paper a data-driven method for the automatic induction of pronunciation rules by means of the analysis of an orthographically transcribed speech corpus is presented. One part of the rules are positive rules describing a transformation that can be applied to the reference transcription of a word in order to produce an alternative pronunciation of that word. Most of the rules however are negative rules describing a transformation that should not be applied in a particular context. The rule learning process first compiles a list of candidate pronunciation rules. A rule pruning procedure, informed by statistics derived from the training corpus, subsequently reduces the rule set without introducing any significant loss of information. Finally, the rules are labeled as either positive or negative. By applying the rules, a consistent set of pronunciation variants of each word is generated. Experiments show that the introduction of such variants in a segment-based recognizer significantly improves the recognition accuracy.
Bibliographic reference. Cremelie, Nick / Martens, Jean-Pierre (1998): "In search of pronunciation rules", In MPV-1998, 23-28.