Evaluating human neural envelope coding as the basis of speech intelligibility in noise

Abstract

Models of speech intelligibility that accurately reflect human listening performance across a broad range of background-noise conditions are clinically important (e.g., for deriving hearing-aid prescriptions, and optimizing cochlear-implant signal processing). A leading hypothesis in the field is that internal representations of envelope information ultimately determine intelligibility. However, this hypothesis has not been tested neurophysiologically. Here, we address this gap by combining human electroencephalography (EEG) with simultaneous perceptual intelligibility measurements. First, we derive a neural envelope-coding metric (ENVneural) from EEG responses to speech in multiple levels of stationary noise, and identify a mapping between the neural metric and corresponding speech intelligibility. Then, using the same mapping, we use only EEG measurements to test whether ENVneural is predictive of speech intelligibility in novel background-noise conditions and in the presence of linear and non-linear distortions. Preliminary results suggest that neural envelope coding can predict speech intelligibility to varying degrees for different realistic listening conditions. These results inform modeling approaches based on neural coding of envelopes, and may lead to the future development of physiological assays for characterizing individual differences in speech-in-noise perceptual abilities.