Human speech learning is remarkably versatile. We are able to recognize and imitate new words after exposure to just one or a few examples, correct how we speak according to feedback, and learn new languages without catastrophically forgetting the languages we have previously learned. To perform the complex speech learning tasks that come so naturally to us in adulthood, we must understand what our representations of speech look like in the first place, and how the same representations can be versatile enough to be used with a variety of contexts and examples. To this end, we propose a computational model of how humans learn to speak from scratch, starting with just the basic components available to an infant: the ability to produce and perceive sounds. Given a perceived speech utterance, the system learns to invert the sound into its underlying articulatory latents, compose primitive articulatory motor programs to create an arbitrarily complex utterance, and update its representation by listening to itself and adapting with a reinforcement learning paradigm.

I’ve enjoyed conducting speech research with Professor Josh Tenenbaum’s group for the past year. I’m excited to continue exploring how humans learn and building on these insights to engineer cognitively plausible models for artificial intelligence.