Search

One of the earliest explanations for good speech intelligibility in poor listening situations was context [Miller et al., J. Exp. Psychol. 41 (1951)]. Context presumably allows listeners to group and predict speech appropriately and is known as a top‐down listening strategy. Amplitude comodulation is another mechanism that has been shown to improve sentence intelligibility. Amplitude comodulation provides acoustic grouping information without changing the linguistic content of the desired signal [Carrell and Opie, Percept. Psychophys. 52 (1992); Hu and Wang, Proceedings of ICASSP‐02 (2002)] and is considered a bottom‐up process. The present experiment investigated how amplitude comodulation and semantic information combined to improve speech intelligibility. Sentences with high‐ and low‐predictability word sequences [Boothroyd and Nittrouer, J. Acoust. Soc. Am. 84 (1988)] were constructed in two different formats: time‐varying sinusoidal sentences (TVS) and reduced‐channel sentences (RC). The stimuli were chosen because they minimally represent the traditionally defined speech cues and therefore emphasized the importance of the high‐level context effects and low‐level acoustic grouping cues. Results indicated that semantic information did not influence intelligibility levels of TVS and RC sentences. In addition amplitude modulation aided listeners’ intelligibility scores in the TVS condition but hindered listeners’ intelligibility scores in the RC condition.

One of the differences in the way a cochlear implant, versus a normally functioning cochlea, delivers sound to auditory centers in the brain is that the implant delivers far fewer channels of frequency information. This experiment investigated how speech intelligibility in cochlear implant (CI) listeners may be related to reduced frequency resolution as opposed to other factors. A closed‐response set Modified Rhyme Test was administered to CI listeners [House et al., J. Acoust. Soc. Am. 37 (1965)] using naturally produced words spoken in isolation. The stimuli presented to the listeners without implants were a reduced‐channel (RC) version of exactly the same stimuli, which were filtered and processed in software in a manner similar to the hardware method described by Shannon [Science 270 (1995)]. The effect of the signal processing was to divide each word into four frequency channels for the first group of RC listeners and into eight channels for the second group. Results indicated final phoneme identific...