Lecture

Computational modeling of non-native phonetic learning and spoken word processing

• dr. Yevgen Matusevych (University of Edinburgh)

Date

Tuesday 10 January 2023

Time

11:00 - 12:00 hour

Address

Snellius
Niels Bohrweg 1
2333 CA Leiden

Room

405

Or attend via Zoom

Zoom link

With the recent rise of deep neural networks for machine learning, these types of models are increasingly being applied to human language acquisition and processing. Interestingly, much less computational modeling work has been conducted to study these processes in bilingual or non-native speakers. In this talk, I present two case studies on native and non-native phonetic learning and spoken word processing in infants and adults. I show how neural networks trained on unsegmented speech data can help us to evaluate and scrutinize linguistic theories in this domain. At the same time, I also highlight potential difficulties associated with the use of such models, in particular their lack of explicit discrete representations.

In the first study, I consider cross-linguistic data from infants' phone discrimination to test several models. The results show that some of them make better models of infant phonetic learning, but also reveal that the existing data may be compatible with more than one theory. In the second study, I test one of the more successful models on data from non-native adult speakers and show that this model can correctly predict behavioral patterns commonly explained by lexical phonology, even though the model is not equipped with explicit knowledge about lexical phonology. I argue that it may be difficult to make clear-cut conclusions about the model’s representations and discuss implications for linguistic theories and computational modeling practices.

• lezing bij science