Comparing accounts of formant normalization against US English listeners' vowel perception

Abstract

Human speech recognition tends to be robust, despite substantial cross-talker variability. Believed to be critical to this ability are auditory normalization mechanisms whereby listeners adapt to individual differences in vocal tract physiology. This study investigates the computations involved in such normalization. Two 8-way alternative forced-choice experiments assessed L1 listeners' categorizations across the entire US English vowel space—both for unaltered and synthesized stimuli. Listeners' responses in these experiments were compared against the predictions of 20 influential normalization accounts that differ starkly in the inference and memory capacities they imply for speech perception. This includes variants of estimation-free transformations into psycho-acoustic spaces, intrinsic normalizations relative to concurrent acoustic properties, and extrinsic normalizations relative to talker-specific statistics. Listeners' responses were best explained by extrinsic normalization, suggesting that listeners learn and store distributional properties of talkers' speech. Specifically, computationally simple (single-parameter) extrinsic normalization best fit listeners' responses. This simple extrinsic normalization also clearly outperformed Lobanov normalization—a computationally more complex account that remains popular in research on phonetics and phonology, sociolinguistics, typology, and language acquisition.