I introduce hierarchical Bayesian modelling and illustrate its relation to traditional ideas in language acquisition. I show how a Bayesian model of phoneme learning can be constructed to solve the problem of learning in the presence of phonetic interference from adjacent segments or speaker-specific variability. It outperforms comparable models based on traditional ideas and is robust to various kinds of (interesting) noise. I then describe an experiment (currently running) which tests the predictions of this model in incomplete-neutralization-type situations.