Modeling talker- and listener-based sources of variability in babble-induced consonant confusions

Speech communication often occurs in the presence of noise. Patterns of perceptual errors induced by background noise are influenced by properties of the listener and of the noise and target speech. The present study introduces a modification of multilevel general recognition theory in which talker- and listener-based variability in confusion patterns are modeled as global or dimension-specific scaling of shared, group-level perceptual distributions. Listener-specific perceptual correlations and response bias are also modeled as random variables. This model is applied to identification-confusion data from 11 listeners' identifications of ten tokens of each of four consonant categories-[t], [d], [s], [z]-produced by 20 talkers in CV syllables and masked by 10-talker babble. The results indicate that dimension-specific scaling for both listeners and talkers provides a good account of confusion patterns. These findings are discussed in relation to other recent research showing substantial listener-, talker-, and token-based sources of variability in noise-masked speech perception.