|Abstract:||Phonological grammars are clearly grounded in phonetics. Within OT, many individual constraints appear to be motivated by particular phonetic facts. The precise connection between such a constraint and its phonetic motivation, however, is not well understood. This paper offers a proposal for how learners might consistently induce perceptually grounded constraints from their acoustic and perceptual experience. The empirical focus of the paper is the constraint *#P 'No word-initial unaspirated p', which is shown to be active in languages including Cajonos Zapotec and Ibibio. Experimental data shows that initial P is uniquely difficult to perceive as a result of its acoustic similarity to initial B. The paper describes a computational model in which virtual learners induce *#P from precisely these phonetic facts. This model achieves realistic perception of acoustically realistic segments; the learner's knowledge of these segments' relative perceptibility then forms the basis for constraint induction.