We proposes a Bayesian account of asymmetries in speech perception: In many languages, listeners show greater sensitivity if a non-coronal sound (/b/, /p/, /g/, /k/) is changed to coronal sounds (/d/, /t/) than vice versa. These asymmetries is predominantly believed to reflect innate constraints from Universal Grammar. Alternatively, we propose that the asymmetries could arise from optimal inference given the statistical properties of different speech categories. In the framework of Bayesian inference, we examined two statistical parameters of coronal and non-coronal sounds: frequencies of occurrence and variance in articulation. In languages where the asymmetries are found, coronal are more frequent and/or more variable than non-coronals. Such differences makes an ideal observer more likely to perceive a non-coronal signal as a coronal segment than vice versa. Without assuming innateness and/or linguistics-specificity, we explain the perceptual asymmetries as a consequence of probabilistic inference similar to asymmetries observed in many other cognitive domains.