Duration: 1:00:49

142 views

5

Speech recognition and understanding technologies rely on supervised learning techniques which typically require tens or hundreds of hours of good quality, human annotated speech in order to train acoustic and language models. In this talk, I argue that it is worthwhile considering an alternative approach, based on unsupervised algorithms, and grounded on the study of human infant language learning. Indeed, during their first years of life, infants spontaneously construct acoustic, language and world models despite large variability in signal quality and amount of parental oversight, across widely different cultures and environments. Reverse engineering this process could therefore enable the development of very robust, flexible and autonomous learning systems as well as enable the modeling and monitoring of normal and impaired language development. I illustrate this approach with results from the recent Zero Ressource Speech Challenge (InterSpeech 2015), and present the Big Baby Data project, a project aimed at constructing a large dataset of parent-infant interactions using kinect sensors.