The goal of this study is to develop a psycho-computational model of human phoneme acquisition that includes the knowledge of linguistic universals to "teach" Artificial Neural Nets incrementally. Long Short-Term Memory (LSTM) artificial neural networks are capable to outperform previous recurrent networks on many tasks ranging from grammar recognition to speech and robot control. Together with our psycho-computational model they are supposed to recognize phonetic features in a way similar to humans learning to understand their first language.