ABSTRACT
Although listeners have knowledge about both categorical and probabilistic phonological patterns, it is not clear if knowledge of probabilistic patterns affects speech processing. One such pattern in English is that words ending in /i/ (e.g. rémedy) are more likely to have antepenultimate stress than penultimate stress (e.g. spaghétti). In the current study, participants extended this trend to ratings of novel words (e.g. bakati). Further, ERPs revealed that real words that violate this trend elicit an early negativity 280–380 ms after onset compared to words that observe this trend. These results indicate that probabilistic phonology interacts with lexical access. More specifically, they suggest that extra processing power is needed to recognise the stress pattern of trend-violating words because of competition between the expectations of the phonological grammar and lexical encoding of trend-violating patterns.
Acknowledgments
Special thanks are due to Lyn Frazier and Joe Pater for contributions to the design of the experiment, as well as to Jason Overfelt for recording the stimuli, and Austin Kopak, Hala Hasan, Maggie Ugolini, Ching Tiv, Elkhansa Elguenaoui, and Kevin Harrington for help collecting data. Thanks are also due to members of the UMass Neurocognition and Perception lab and Sound Seminar, audiences at the 22nd Manchester Phonology Meeting and LabPhon14, as well as two anonymous reviewers for useful discussion and comments.
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
1. Whenever the first consonant is not nasal.
2. Consider long words such as “àbracadábra”, or “apòtheósis”. Words of this length can have a secondary stress on the first syllable, but never a primary stress. “ábracadàbra” would not be a possible English word.
3. Many affixes in English affect stress placement – for example, the affix “-ity” requires that stress be antepenultimate, as can be seen in the pair sólemn ∼solémnity.
4. Thanks to Adam Albright for sharing his implementation of the GNM model.
5. These mean ratings were roughly normally distributed for the nonwords, but not for real words, whose ratings were clustered at the high end of the scale.
6. Due to a coding error, six participants did hear some of the nonwords exactly twice throughout the course of the experiment.