Flexible neural encoding predicts the comprehension of degraded speech

Abstract

How listeners track a variable and continuous acoustic speech signal and parse it into meaningful linguistic representations is a question central to cognitive neuroscience. Moreover, the resilience of this process to acoustic signal degradation is not fully understood. The current study consists of a listening task wherein participants (n = 38) were presented with a naturalistic story whilst undergoing continuous electroencephalography (EEG). Critically, we manipulated access to speech information over two independent dimensions: Spectral clarity, which ranged from unprocessed to severely spectrally degraded; and language, which was either English, spoken by the participants, or Dutch, an incomprehensible language that is closely related to English. All stimuli were produced by the same bilingual speaker. We applied banded regression to model the neural response to a set of acoustic and linguistically derived features. We found that there is no single speech feature for which neural encoding reliably predicted speech understanding in individual listeners or conditions. Yet, differences in feature weights between experimental conditions can be combined into a composite score that strongly predicted individual subjective comprehension of spectrally degraded but linguistically accessible speech in data from held-out participant. Hence, the manner in which neural encoding flexibly adapts to listening context is associated with advantageous perceptual strategies for degraded speech. These findings underscore the complex interplay between low- and high-level features during speech listening and illustrate inter-condition differences in the neural encoding of these properties. Our results highlight natural variation within and between individuals in the encoding of acoustic and linguistic features which provides a potential pathway towards individualised assessment of clinical populations.