Dealing with unexpected words in automatic recognition of speech

Abstract

Unexpected words attract listener's attention. They are information-rich and getting them right is important for human communication. In the automatic recognition of speech (ASR), words that are not in the expected lexicon of the machine are typically substituted by some acoustically similar but nevertheless wrong words. The article discusses reasons for this undesirable behavior of the machine, describes some known examples of dealing with the unexpected words in human speech perception and their implications, and proposes an alternative architecture of ASR that could alleviate some of the problems with the unexpected acoustic inputs. Some published experimental results from using this alternative architecture are given. © 2011 Springer-Verlag.