Miller and Nicely’s confusion data are predicted by Fletcher’s AI

Abstract

Starting from the confusion matrix data Pij(SNR) of Miller and Nicely, where i indexes the stimulus and j the response, the average performance intensity (PI) function, over the 16 consonants, may be defined as P̄(SNR)=∑i=116Pii(SNR). These data are for five female talkers. The average speech power for five female talkers is known from Dunn and White. Thus, the articulation index (AI) may be computed at each SNR value (the maximum AI value is 0.6). This allows one to re−express P̄(SNR) as P̄(AI). When the resulting function is compared to Fletcher′s formula for the AI, P(AI)=1-eminAI, with emin=0.015, the agreement is nearly perfect. Thus, the Miller and Nicely average phone data may be modeled by AI theory. This result seems astounding, as it falls outside normal realm of the AI, which was intended for a much larger mix of sounds, not 16 consonants and a fixed vowel. Most individual consonant PI functions obey the same Fletcher AI formula, but with different values of emin.