Word models representing single pronunciations are often too simple for continuous speech recognition tasks. Multiple pronunciation word models are likely to improve the recognition performance. In this paper a method is introduced to deal with pronunciation variations originating from articulatory interactions between words. These variations are described by rewrite rules which are inferred automatically from a training set. Test results show that a significant reduction of the word error rate can be obtained with the described method.
Bibliographic reference. Cremelie, Nick / Martens, Jean-Pierre (1995): "On the use of pronunciation rules for improved word recognition", In EUROSPEECH-1995, 1747-1750.