This PhD is funded by the French ANR, under a 4 years' project on Sensorimotor integration of variability during birdsong learning. The applicant will develop an artificial neural model, developmental and brain-inspired, to learn the sound structure in real time and without explicit supervision. Until now, AI models for developmental learning of vocalizations have been solely validated by comparison against a human-annotated corpus and not yet via direct sensorimotor interactions with living animals. We expect to do so with an interactive robot under the framework of active inference and predictive coding.