MA thesis defense: Lexical competition and predictive certainty in speech recognition: Modulations of pre-activation negativity amplitude by continuations, entropy and inhibition

17 January 2022 10:15 to 12:00 Seminar

Anna Hjortdal

Opponent Abdolnoor Khaleghi

Examiner Marianne Gullberg


A growing body of evidence suggests that speech recognition is facilitated by rapid activation of possible lexical candidates and subsequent competition and selection. An event-related potential (ERP) component 136-204 ms after word onset, the pre-activation negativity (PrAN), correlates with lexical competition. Further, the increased negativity for low lexical competition correlates with increased activity in Broca’s area. The effect has been interpreted as reflecting predictive certainty about the unfolding word and inhibition of irrelevant candidates. The present study investigated effects of lexical competition, predictive certainty and inhibition when only the first 2-3 speech sounds of words were available to listeners. The measures were calculated from a combined pronunciation lexicon and frequency list with a Python script. Correlations with ERP data from two experiments investigating Danish language were explored. In line with previous findings, word beginnings with few continuations (low lexical competition) showed more negative ERPs than word beginnings with many continuations. For one experiment, there was an inverse correlation between the number of possible continuations and PrAN amplitudes. This is consistent with parallel activation of multiple lexical candidates as predicted by a number of models of speech perception and spoken word recognition. According to the distributed model of speech perception, there is an inverse relationship between the number of pre-activated words and their lexical activation. The increased PrAN for low competition might reflect stronger lexical activation of one or a few candidates. Word beginnings with low entropy (more certainty) also showed more negative amplitudes. Within the framework of the predictive coding model of speech recognition, the findings suggest an interplay between pre-activated lexical candidates and updated phonetic expectations at lower-level, primary processing areas. There were no effects of inhibition of irrelevant candidates immediately after word onset. When more of the speech signal became available, prosodic cues during or immediately after the stressed vowel which inhibited many candidates yielded more negative, right frontal effects than speech sounds inhibiting few candidates. One interpretation is that inhibition only started later, as more of the word unfolded.


About the event:

17 January 2022 10:15 to 12:00



Save the event to your calendar