The production and perception of word boundaries

The recognition of continuous speech presents listeners (human or machine) with a problem which does not arise in the recognition of isolated words, and does not confront the reader in most orthographies. The act of recognition is the identification of an input as something we already know; what we already know is not the whole of an input utterance, because human memory is not infinite, and it would be impossible to store in our memories every complete utterance we might ever hear. Therefore the entries in our mental lexicon must be discrete, and recognition will involve finding these discrete lexical units as sound patterns in the speech signal input, and matching them to lexical entries in order to determine their meaning. The problem for listeners arises in the fact that speech is continuous: lexical unit boundaries are not reliably marked. Finding the boundaries between such units — segmenting the speech signal — is therefore a nontrivial task for all listeners.