Characterizing neural mechanisms of attention-driven speech processing

The ability to selectively attend to speech in adverse listening environments plays a key role in human communication. However, listening to speech in everyday multi-talker environments can be challenging for hearing-impaired listeners,even when wearing modern hearing aids. Studies in normal-hearing listeners have shown that it is possible to decode which of two speakers a listener is attending to from scalp electroencephalography (EEG).

This has led to the idea that EEG-based brain-computer interfaces (BCIs) could be integrated into hearing assistive devices to help hearing-impaired listeners by selectively amplifying attended sound sources. To accomplish this, however, single-trial EEG measures of attention to speech must first be investigated in hearing-impaired listeners, in everyday listening scenarios, and under different task demands.

This thesis explored single-trial cortical EEG correlates of selective attention to speech. In the first study, the influence of sensorineural hearing loss on cortical EEG responses to tones and to naturalistic speech was investigated. It was shown that hearing impairment enhances the fidelity of the low-frequency cortical EEG entrainment to envelopes of simple tones and to envelopes of attended speech streams.

For loudness-matched competing speech streams, the attended target could be classified from single-trial EEG responses with equally high classification accuracies in normal-and hearing-impaired listeners. The second study explored single-trial EEG correlates of selective auditory attention to speech in reverberant, multi-talker environments.

It was shown that the attentional selection of normal-hearing listeners could be decoded from single-trial EEG data with equally high classification accuracies in both anechoic-and reverberant listening environments. The third study analyzed how different constraints on EEG-based stimulus-response models influence the predictive power of the models.

The results from this study suggested that stimulus-response model regularization is important for maintaining high classification accuracies with backward decoding models. The fourth study investigated if working memory demands affect single-trial EEG correlates of speech envelope processing. Using an auditory n-back task, it was found that working memory load can affect spatio-spectral EEG power in the theta and alpha bands and that EEG measures of speech envelope entrainment decrease with high taskload.

The fifth study investigated whether EEG-based attention decoding can be achieved in a real-time closed-loop BCI system. Here, it was shown that a hearing-impaired listener was able to selectively amplify attended speech using a closed-loop EEG-based attention decoding BCI system. Overall, the work presented in this thesis suggests that EEG-based attention decoding may have relevance for future BCI systems.