Investigating the perception of noise-vocoded speech - an individual differences approach

Abstract

We used a cochlear implant simulation (noise-vocoded speech) to investigate speech recognition and perceptual learning in normal-hearing adult speakers of English. In two separate sessions (1-2 weeks apart), 28 listeners were tested on recognition of noise-vocoded Sentences, Words, and isolated segments (Consonants and Vowels). There was evidence of significant perceptual learning that survived until Session 2 for all tasks. An individual differences analysis of Session 1 data suggested two independently-varying 'levels' of processing at work in the initial perception of the distorted speech stimuli - a 'top-down' listening mode making use of contextual and lexical information, and a 'bottom-up' mode focussed on acoustic-phonetic discriminations. By Session 2, a more generalised listening mode emerged, reflecting listeners' consolidation of basic sound-to-representation mappings. Further exploration of Consonant and Vowel confusion data (using Information Transfer analyses) suggested that better speech recognition performance may be achieved through more efficient use of the preserved cues to duration and voicing in noise-vocoded stimuli, but that listeners failed to take full advantage of such information. We conclude that training regimes involving directed attention to specific features, such as vowel length, may help to improve performance with noise-vocoded speech.