Prediction from Statistical Learning Aids Auditory Scene Analysis.

Statistical regularities support auditory scene analysis across multiple levels. While acoustic regularities like comodulation and harmonicity aid bottom-up perceptual grouping, higher-level regularities like linguistic or musical structure must be learned to form a mental "schema" of statistical patterns. Although learned schemas may benefit comprehension by helping listeners perceptually separate and/or attend to a target sound stream in an acoustic mixture, the underlying mechanisms are unclear. Here, we used a statistical learning paradigm to expose listeners to sequences of speech syllables with fixed transitional probabilities, forming an artificial "language" of trisyllabic words. Following exposure, participants attended to one of two concurrent syllable streams and detected target syllables. Detection performance improved when the attended stream conformed to the statistical structure learned implicitly during exposure, with a larger benefit in the presence of a competing stream than in quiet. In contrast, predictability of the unattended stream had no effect on performance. Electroencephalography revealed that predictable targets elicited earlier parietal P300 "target-recognition" responses and enhanced neural tracking of the attended stream, with additional signatures of predictive processing observed even in the absence of targets. These findings demonstrate that learned statistical regularities enhance listening in noise by enabling predictive, schema-based selection of relevant input. Rather than facilitating automatic segregation of competing sounds, learned lexical schemas support auditory scene analysis through attentional template matching. Our findings establish a direct mechanistic link to the role of prediction in schema-based listening in noise.