Abstract:

Articulation theory is a computational model designed to predict speech recognition scores for nonideal listening channels. Fletcher's original formulation [H. Fletcher and R. H. Galt, J. Acoust. Soc. Am. 22, 89--151 (1950)] has been overlooked, presumably due to its complexity and the availability of simplified versions such as ANSI S3.5-1969. The Fletcher and ANSI versions were compared on their ability to predict articulation scores using data collected by Rankovic et al. [J. Acoust. Soc. Am. 90, 2319(A) (1991); 91, 354--362 (1992)]. In those studies, consonant-correct scores were obtained from five normal-hearing and seven hearing-impaired subjects in the presence of octave-band noises centered on 0.5, 1, or 2 kHz and noise-masking patterns were obtained from all subjects in all conditions. The speech was amplified for the hearing-impaired subjects. For normal-hearing subjects, both models made good predictions of consonant-correct scores. For hearing-impaired subjects, the ANSI version overestimated scores for conditions with articulation indices (AIs) greater than about 0.5, whereas Fletcher AIs were smaller and predictions more accurate. Differences in the AI calculations that account for these results will be discussed. [Work supported by NIH.]