Learning from a tutor: embodied speech acquisition and imitation learning

摘要

This work presents a new developmentally inspired data-driven framework to bootstrap speech perception and imitation abilities in interaction with a tutor. The proposed system architecture extends our work presented in [1], that implements a cascade of interconnected layers to acquire the structure of speech in terms of phones, syllables and words. Here, we show how to couple such a perceptual model with a speech imitation system that is based on an acoustic synthesizer bound to produce speech sounds with a child's voice. We present an interaction scheme to learn the correspondence model between the tutor's and the system's voice. It acquires such a model by linking its own phonemic primitives to the imitating feedback from its tutor. The model can then be used to map a tutor's utterance into a continuous activity in the system's motor space. Furthermore, we propose how such motor activities can be used to imitate words. Finally, we embed this architecture into an embodied autonomous learning and interaction system to provide a grounding for the speech models to be acquired and a perceptual input to trigger speech production.