Is the sine-wave speech cocktail party worth attending?

Abstract

Listeners are remarkably adept at recognising speech that has undergone extensive spectral reduction. Natural speech can be reproduced using as few as three time-varying sinusoids mimicking the corresponding speech formants. Untrained listeners are able to transcribe this `sine-wave' speech with a high degree of reliability. Coherent phonetic percepts generated by sine-wave speech occur despite an apparent lack of those cues on which low level auditory grouping processes are believed to operate. Consequently, it has been proposed that speech perception is governed by processes operating independently of those described by auditory scene analysis. This paper re-examines the evidence provided by previous perceptual studies of sine-wave speech and presents new data, from perceptual studies employing stimuli constructed from simultaneous sine-wave wave speech sources. These new studies suggest that in conditions that are closer to those of everyday listening, grouping cues have an important role in the formation of coherent speech percepts. In conjunction with these perceptual studies, results from automatic segregation and recognition tasks suggest that sine-wave speech contains sufficient low level, non speech-specific, structure to allow partial descriptions of sine-wave sources to be recovered from two-source mixtures. It is argued that these partial descriptions may be sufficient to account for the limited intelligibility observed in two-source sine-wave speech listening tests.