Regressive spectral assimilation bias in speech perception

Speech perception presents a parsing problem: construing information from the acoustic input we receive as evidence for the speech sounds we recognize as language. Most work on segmental perception has focused on how listeners use differences between successive speech sounds to solve this problem. Prominent models either assume (a) that listeners attribute acoustics to the sounds whose articulation created them, or (b) that the auditory system exaggerates the changes in the auditory quality of the incoming speech signal. Both approaches predict contrast effects in that listeners will usually judge two successive phones to be distinct from each other. Few studies have examined cases in which listeners hear two sounds in a row as similar, apparently failing to differentiate them. We examine such under-studied cases. In a series of experiments, listeners were faced with ambiguity about the identity of the first of two successive phones. Listeners consistently heard the first sound as spectrally similar to the second sound in a manner suggesting that they construed the transitions between the two as evidence about the identity of the first. In these and previously reported studies, they seemed to default to this construal when the signal was not sufficiently informative for them to do otherwise. These effects go unaccounted for in the two prominent models of speech perception, but they parallel known domain-general effects in perceptual processing, and as such are likely a consequence of the structure of the human auditory system.