Speaker: Jonathan Howell (McGill)
When: Friday, March 16 at 3:30 pm
Where: Leacock Building, room 232
Title: Acoustic classification of contrastive focus: On the web and in the lab

In the study of meaning and prosody, the ideal investigation requires data which are at once context-rich, context-controlled, quantitative and reproducible. On their own, the familiar methodologies of personal introspection, laboratory experimentation and corpus analysis typically fail on one or more on these accounts. In this talk, I describe a method for studying the acoustic realization and semantic/pragmatic conditioning of prosody, which combines naturallyoccurring speech harvested online and elicited speech recorded in the laboratory. I examine the English comparative construction: e.g. than I did in (1) and (2).

(1) She liked the song more than [I]F did.
(2) I wish I had done more than I [did]F

Linguistic theory predicts the location of contrastive focus according to (lack of) co-reference in the antecedent clause. I test this claim by training machine learning classifiers (SVMs and LDA) to predict semantic focus using only acoustic parameters. By training the classifier on naturallyoccurring web data and testing it on controlled lab data (and vice versa), it is possible to crossvalidate the two. Because the method allows one to identify which acoustic cues reliably encode contrastive focus across contexts and speakers, we are also able to engage with two longstanding debates in phonology: (i) the relationship between relational and paradigmatic comparison and (ii) the relationship between prominence, pitch and stress. Finally, I also report the results of a perception experiment which asks whether the machine learning classification makes use of the same acoustic cues as human listeners. This work is part of a project “Harvesting Speech Datasets for Linguistic Research on the Web” with collaborators Mats Rooth (Cornell University) and Michael Wagner (McGill University).