Speech Science Forum -- Alexis Deighton MacIntyre

During speech listening, patterns of neural activity become temporally coupled to stimulus features, which can be measured using electroencephalography (EEG) and other neuroimaging techniques. This opportunity to "decode" from the listening brain may provide important insights into the nature of speech perception. It also holds promise as a candidate objective measure to guide fitting and assess outcomes for cochlear implant (CI) listeners. The effects of CI-associated spectral degradation on neural speech decoding, however, remain unclear. In particular, the interplay between speech intelligibility and acoustic signal quality is debated, with contradictory evidence for and against the enhancement of neural speech tracking for comprehended speech. Importantly, few studies independently manipulate acoustic properties and intelligibility. In the current experiment, we simulate CI listening by presenting natural and spectrally degraded speech to typically hearing listeners (n = 38) undergoing EEG recording. To dissociate acoustic-sensory processing from comprehension, we use intelligible (English) and non-intelligible (Dutch) speech produced by one bilingual speaker. We promote auditory attention throughout the experiment using a novel prosodic target detection task that is not contingent on speech understanding. Decoding models were then trained to reconstruct speech parameters from held-out neural response data, with correlations between reconstructed and true stimulus features providing a measure of neural speech decoding. In this talk, I will discuss our first analysis of this data set, which focused on the speech amplitude envelope, a commonly used acoustic feature within the speech decoding literature. Turning to our ongoing work, I will share early results concerning how well other, less established acoustic properties of the speech signal can be reconstructed from EEG. Finally, I compare acoustic to linguistic feature decoding with a view towards identifying the parameters that best indicate differences in speech processing across listening conditions.

zoom link: https://ucl.zoom.us/j/92052680901