5th European Conference on Speech Communication and Technology

Rhodes, Greece
September 22-25, 1997

Statistical Techniques for Robust ASR: Review and Perspectives

Jerome Bellegarda

Apple Computer, USA

Speech recognition performance degrades significantly when a mismatch occurs between training and operating conditions. To reduce this mismatch, it is often necessary to characterize the mapping between two environments. A number of statistical approaches have been developed for this purpose. They can be classified as either predictive or adaptive, depending on what information is available regarding the operating environment. This paper reviews a selected subset from both categories, and discusses possible future directions of improvement.

Full Paper

Bibliographic reference.  Bellegarda, Jerome (1997): "Statistical techniques for robust ASR: review and perspectives", In EUROSPEECH-1997, KN33-KN36.