Representation and Process in Human Speech Comprehension

The relationship between human speech processing and Computer speech processing has been a historically difficult one; a series of failed marriages ending in divorce. At the moment, we are in a period of divorce, where Statistical engineering methods (Jelinek, 1976), based around HMM’s and their relatives, dominate both industrial and academic approaches to speech processing by machine. The reason for this is simply that knowledge-based1 approaches have not been successful in offering acceptable levels of performance. Only Statistical methods have seemed to offer a stable and incremental basis for achieving the goal of speaker-independent continuous speech recognition for large vocabularies.