Effects of linguistic context and noise type on speech comprehension

Abstract

Introduction: Understanding speech in background noise is an effortful endeavor. When acoustic challenges arise, linguistic context may help us fill in perceptual gaps. However, more knowledge is needed regarding how different types of background noise affect our ability to construct meaning from perceptually complex speech input. Additionally, there is limited evidence regarding whether perceptual complexity (e.g., informational masking) and linguistic complexity (e.g., occurrence of contextually incongruous words) interact during processing of speech material that is longer and more complex than a single sentence. Our first research objective was to determine whether comprehension of spoken sentence pairs is impacted by the informational masking from a speech masker. Our second objective was to identify whether there is an interaction between perceptual and linguistic complexity during speech processing.

Methods: We used multiple measures including comprehension accuracy, reaction time, and processing effort (as indicated by task-evoked pupil response), making comparisons across three different levels of linguistic complexity in two different noise conditions. Context conditions varied by final word, with each sentence pair ending with an expected exemplar (EE), within-category violation (WV), or between-category violation (BV). Forty young adults with typical hearing performed a speech comprehension in noise task over three visits. Each participant heard sentence pairs presented in either multi-talker babble or spectrally shaped steady-state noise (SSN), with the same noise condition across all three visits.

Results: We observed an effect of context but not noise on accuracy. Further, we observed an interaction of noise and context in peak pupil dilation data. Specifically, the context effect was modulated by noise type: context facilitated processing only in the more perceptually complex babble noise condition.

Discussion: These findings suggest that when perceptual complexity arises, listeners make use of the linguistic context to facilitate comprehension of speech obscured by background noise. Our results extend existing accounts of speech processing in noise by demonstrating how perceptual and linguistic complexity affect our ability to engage in higher-level processes, such as construction of meaning from speech segments that are longer than a single sentence.

Keywords: acoustic challenges; linguistic complexity; linguistic context; perceptual complexity; speech comprehension; speech in noise; speech perception; task-evoked pupil response.

Figure 1

Aggregated pupil dilation relative to baseline (in mm) across the time series (in ms) of stimulus presentation and the post-stimulus retention period prior to comprehension question presentation. Pupil dilation data measured in the babble noise (Bab) condition are pictured in gray, while pupil dilation measured in the Steady-State Noise (SSN) condition are in yellow. Within noise conditions, each context condition is represented by a different shape. The window of time in which pupil dilation data were aggregated for analysis is labeled. Vertical black lines mark sentence pair mean onset and offset. Error bars are ±1 standard error.

Figure 2

Behavioral data of comprehension accuracy and reaction time. Panel (A) depicts the proportion of comprehension questions answered correctly for each context condition (expected exemplar, EE; within-category violation, WV; between-category violation, BV) in the babble noise (Bab) condition and the Steady-State Noise (SSN) condition. Panel (B) depicts how much time (in seconds) elapsed before participants provided an answer to post-stimulus comprehension questions for each context condition (EE, WV, BV) in both noise conditions (babble and SSN). The boxes indicate first quartiles, third quartiles, and medians; the whiskers indicate maximum and minimum.

Figure 3

Mean peak pupil dilation measured within the defined window of interest (−1000 to +1500 ms relative to speech offset) for both noise conditions by context condition. Individual participant data is overlayed. Error bars are ±1 standard error.