Perception of coarticulated speech with contrastively enhanced spectrotemporal patterns

Abstract

High-level contrastive mechanisms cause perception of auditory events to be influenced by spectral and temporal properties of surrounding acoustic context, and may play a role in perceptual compensation for coarticulation in human speech. However, it is unknown whether auditory contrast is incorporated optimally to compensate for different speakers, languages and situations or whether amplification of the processes involved would provide additional benefit, for example, in the perception of hypoarticulated speech, under adverse listening conditions, or in an incompletely acquired language. This study examines effects of artificial contrastive modification of spectrotemporal trajectories on the intelligibility of connected speech in noise by native and non-native listeners. Adopting methods known to improve automatic classification of speech sounds, we model contrast-providing context as an averaged estimated vocal tract function (LPC-derived log area ratio coefficient vector) over a Gaussian-weighted temporal window. Local coefficient values are adjusted from this context based on previously observed contrastive perceptual tendencies, and the intelligibility of the resulting speech is compared with that of unmodified trajectories across listener language backgrounds. Results are discussed with respect to implementation and applicability of general auditory processes.