Expectation-based decoding of ambiguous speech

Introduction: Accounts of predictive coding suggest that the brain continuously anticipates upcoming sensory events in order to reduce ambiguity by integrating incoming information with prior knowledge. The transient nature of spoken utterances should make speech a prime example of a process depending on prior information. We conducted an fMRI study involving a delayed-match-to-sample task with pairs of normal and degraded sentences. Methods: The degraded sentences merely contained the prosodic parameters of the original sentence and were therefore physically incomprehensible. Importantly, they could suddenly be understood when preceded by their original version. Whereas the target sentence was always a degraded sentence that matched the prior sentence in half of all trials, the reference sentence within each pair alternated between normal and degraded sentences. Results: The behavioural data showed similar discrimination performance and speed of the match/non-match decisions irrespective of the type of reference sentence. Assessing a degraded sentence against a preceding normal sentence evoked significantly higher activation in the left inferior frontal gyrus (IFG), left middle temporal gyrus (MTG) and left thalamus. This network was further dissociated into left IFG being selectively activated by precedence of non-matching normal sentences and left MTG and thalamus for matching normal sentences. Conclusions: These results support theories of Bayesian perception which propose that predictions specified at higher cortical levels can be used to decode ambiguous stimuli. When prior information facilitated the comprehension of the degraded sentences the left MTG and the left thalamus were selectively activated, highlighting a role for these areas in meaning extraction. In contrast, selective activation of the left IFG reflected the search for meaningful information in ambiguous speech material that could not be decoded because of mismatches with the prior information.