Learning complex mappings between various modalities (typically articulatory, somato­sensory and auditory) is a central issue in computationally modeling speech acquisition. These mappings are generally non­linear and redundant, involving high dimensional sensorimotor spaces. Classical approaches consider two separate phases: a relatively pre-determined exploration phase analogous to infant babbling followed by an exploitation phase involving higher level communicative motivations. In this paper, we consider the problem as a developmental robotics one, in which an agent actively learns sensorimotor mappings of an articulatory vocal model. More specifically, we show how intrinsic motivations can allow the emergence of efficient exploration strategies, driving the way a learning agent will interact with its environment to collect an adequate learning set.