Learning to Sing like a Bird: The Self-Supervised Acquisition of Birdsong

Abstract

This paper presents a new framework for self-supervised sensorimotor learning. We demonstrate this framework with a system that learns to mimic a zebra finch, directly modeled on the dynamics of how male fledglings acquire birdsong from their fathers. Our system first listens to the song of an adult finch. By listening to its own initially nascent attempts at mimicry through an articulatory synthesizer, the system organizes motor maps generating its vocalizations. Our approach is founded on the notion of cross-modal clustering, introduced in (Coen 2005, 2006a), and is unusual for its recursive reuse of perceptual mechanisms in developing motor control. In this paper, we outline this framework, present its results on the unsupervised acquisition of birdsong, and discuss other potential applications.