Neural Processing of Noise-Vocoded Speech Under Divided Attention: An fMRI-Machine Learning Study

In real-life interaction, we often need to communicate under challenging conditions, such as when speech is acoustically degraded. This issue is compounded by the fact that our attentional resources are often divided when we simultaneously need to engage in other tasks. The interaction between the perception of degraded speech and simultaneously performing additional cognitive tasks is poorly understood. Here, we combined a dual-task paradigm with functional magnetic resonance imaging (fMRI) and machine learning to establish the neural network supporting degraded speech perception under divided attention. We presented 25 human participants with noise-vocoded sentences while they engaged in a concurrent visuomotor recognition task, employing a factorial design that manipulated both speech degradation and task difficulty. Participants listened to eight-band (easier) and four-band (more difficult) noise-vocoded sentences, while the Gabor task featured two difficulty levels, determined by the angular discrepancy of the target. We employed a machine learning algorithm (Extreme Gradient Boosting, XGBoost) to evaluate the set of brain areas that showed activity predicting the difficulty of the speech and dual tasks. The results illustrated intelligibility-related responses in frontal and cingulate cortices and bilateral insulae induced by divided attention. Machine learning further revealed modality-general and specific responses to speech and visual inputs, in a set of frontotemporal regions reported for domain-general cognitive functions such as attentional control, motor function, and performance monitoring. These results suggest that the management of attentional resources during challenging speech perception recruits a bilateral operculo-frontal network also associated with processing acoustically degraded speech.