EEG-based Decoding of Auditory Attention to Conversations with Turn-taking Speakers
ObjectivesAuditory attention decoding (AAD) refers to the process of identifying which sound source a listener is attending to, based on neural recordings, such as electroencephalography (EEG). Most AAD studies use a competing speaker paradigm where two continuously active speech signals are simultaneously presented, in which the participant is instructed to attend to one speaker and ignore the other speaker. However, such a competing two-speaker scenario is uncommon in real life, as speakers typically take turns rather than speaking simultaneously. In this paper, we argue that decoding attention to conversations (rather than individual speakers) is a more relevant paradigm for testing AAD algorithms. In such a conversation-tracking paradigm, the AAD algorithm focusses on switching between entire conversations, resulting in less frequent attention shifts (ignoring turn-taking within conversations), thereby allowing for more relaxed constraints on the decision time.

DesignTo test AAD performance in such a conversation-tracking paradigm, we simulated a challenging restaurant scenario with three simultaneous two-speaker conversations, which were podcasts presented in front of the listener and in the back left and back right of the room. We conducted an EEG experiment on 20 normal-hearing participants to compare the performance of AAD in the commonly used competing speaker paradigm with two speakers versus the conversation tracking paradigm with 2 or 3 conversations, each containing two turn-taking speakers.

ResultsWe found that AAD, using stimulus decoding, worked well under all experimental conditions, and that the accuracy was not influenced by the direction of attention, the proximity to the target conversation, or the presence of within-trial attention switches (versus a condition with sustained attention). Given the challenging scenario, we probed for the participants listening experience and found a correlation between the neural decoding performance and the perceived listening effort and self-reported speech intelligibility. To gain insight into the speech intelligibility of the participants in our setup, they performed a speech-in-noise test (Flemish matrix sentence test), but we did not find a correlation between the speech intelligibility performance and the AAD performance.