Bart De Boer - Speaker

Yannick Jadoul - Contributor

Marnix Van Soom - Contributor

When studying the evolution of speech, an intriguing difference between humans and other primates is humans’ ability to find structure in continuous signals: we are able to split up the continuous, noisy acoustic (or visual) signals we perceive into discrete building blocks, and we are able to learn the rules of how to combine these build- ing blocks into larger utterances. This lies at the basis of our ability use language in an unlimited way. The underlying mechanisms, however, are largely unknown. We apply techniques from artificial intelligence, data mining and machine learning to build cognitively plausible models for investigating candidate mechanisms that underlie these abilities, and for investigating possible scenarios for how they could have evolved from known abilities of other primates. More precisely, at the moment we focus on frequent sequence mining and on probabilistic approaches for analysis of speech as candidate techniques for learning structure in continuous, noisy sig- nals. The sequence mining approach fits into a family of data mining techniques called frequent pattern mining (Aggarwal & Han, 2014). These algorithms are used to extract simple and explainable patterns from a large corpus of data, based on the occurrence frequency of said patterns, but existing methods focus on discrete, symbolic data. Probabilistic approaches to analysis of speech directly model the sta- tistical inference process done by humans to distinguish and recognize the signal’s components (Jaynes 1987; Chater et al. 2010). This approach aims to understand why human speech recognition is so robust and needs so little data compared to computer speech recognition. Although we are aware that these are undoubtedly not precisely the mechanisms that exist in humans, having more insight in these candidate mechanisms will help to better understand what may have happened in the evolution of speech.
4 Mar 2019

Event (Workshop)

TitleISLE Inaugural Workshop
Web address (URL)
LocationUniversity of Zürich
Degree of recognitionInternational event

ID: 44668520