Speech perception as a multilevel processing system

Abstract

This is the report of an experiment that emphasized how variability in voice characteristics will affect a listener's ability to process sentences. Consideration is also given to the description of a possible model for the processing of continuous speech. As part of the experiment, use was made of a number of speakers, each of whom recorded the same set of sentences. The recorded sentences were used in order to form a pair of experimental tapes. On one tape, sentences were arranged randomly such that the listener could predict neither the content of a sentence nor the voice of the speaker who had produced it. On a second tape, sentences were arranged in blocks by speakers such that there was uncertainty as to content but not with respect to voice. Four groups of subjects listened to these tapes, with each tape being processed under a low and high condition of background noise. In interpreting the data, emphasis was placed on the idea that the speech signal contains a variety of characteristics or features and that these features are processed by three interacting subsystems. One is a prosodic system, responsible for the segmentation of continuous speech into sentences, phrases, and words. It attempts to establish a context within which a second system, responsible for the processing of words and syllables, can operate. However, this pair of systems is speaker-dependent in that it makes use of features that need to be adjusted on the basis of an assessment of voice characteristics. Thus, the model also provides for the inclusion of a third system, responsible for the assessment of voice characteristics. In effect, this third subsystem is responsible for “training” the two primary systems in order to normalize for differences in individual voice characteristics. We had predicted that performance would be uniformly more accurate in the blocked condition than in the unblocked condition since one type of uncertainty would have been eliminated. The results confirm this hypothesis only in part since there is an interaction with variations in backgroud noise. However, a plausible explanation is offered for these findings and for a variety of related results.