The University of Arizona

Events & News

Cog Sci Brown Bag Seminar

DateFriday, September 18, 2009
Time12:00 pm
LocationGS 906
SpeakerNatasha Warner
TitleAssociate Professor
AffiliationDept of Linguistics, University of Arizona

Real-life Speech Processing: Combining Big Picture and Detail

ABSTRACT: In the field of Cognitive Science, there is a balance between studying how human cognition works in realistic, daily life situations, and studying small parts of it in controlled situations. Real-life situations tend to be extremely variable, making it difficult to construct careful, controlled experiments. Even if one sets out to study how a particular aspect of the cognitive process works in relatively natural settings, one can use both more and less naturalistic experiments to study it.

My lab is working on reduced speech, such as speakers and listeners use in their daily life conversational interactions. The vast majority of past research on speech has used careful "lab speech," which is very different from the reduced speech we all use for most of our communication. In reduced speech, phrases can lose sounds, syllables, or even entire words, and yet they sound natural and intelligible to listeners, at least with context. For example, we have conversational recordings in which the sequence "do you have time" was pronounced as "dyutem," losing two syllables and the entire word "have." "He was" can sound exactly like "he's" out of context, but clearly like "he was" in context. Even people who have worked extensively with speech are often not aware of how much variability there is in daily-life use of conversational speech.

In studying how humans communicate in daily life, one can pick a single phenomenon to focus on, such as how a particular sound is reduced in conversational speech. However, one eventually feels that this does not get at what one wanted to study in the first place: the massive variability of open conversation that deletes and alters large parts of a phrase. Alternatively one can attempt to study the entire range of variability that occurs in casual conversation, but one then finds that one has so little control over the materials or data that it can be difficult to locate significant effects.

In the current talk, I will discuss a series of experiments my lab has used to study reduced speech by approaching it from both directions, the controlled and the natural. Some of the methods using more natural stimuli have failed because of stimulus variability, but some have also succeeded. The controlled methods have led to reliable results. We combine the methods in order to both obtain usable results and keep the big picture of natural, massive variability in sight. In the current talk, I will present several results from this work, showing which sources of information speakers and listeners integrate in producing and perceiving conversational, reduced speech. The combination of detailed and broad methods may serve as a model for learning about how any aspect of human cognition works in real-life settings