Models of Phonetic Recognition II: an Approach to Feature-Based Recognition

Abstract

An approach to speech recognition is proposed in which phonetic features are identified as acoustic properties in the speech signal, and lexical items are accessed directly without explicitly labeling phonetic segments. A possible advantage of such an approach is that a feature representation shows minimal modification as a consequence of the deletions and assimilation phenomena that occur in natural speech. Problems of determining acoustic correlates of features and of representing lexical items in terms of features are discussed. In this paper I would like to argue that there are advantages to be gained by using phonetic features as primary units for identifying words. I hope to show that variability that occurs from speaker to speaker and from context to context can be taken into account in a natural way if features are used for representing utterances and if they form the building blocks for larger units by means of which utterances are identified. Before discussing some of the advantages of features, and the structure of a speech recognition procedure based on features, let me first review some of the basic ideas underlying the concept of features.