Machine learning approaches to hearing loss estimation can significantly reduce the number of required experiments, but require a good probabilistic hearing loss model. In this work we introduce such a model, obtained by fitting a mixture of Gaussian processes to a vast database containing audiometric records of around 85k people. The learned model can be used as a prior distribution for hearing loss, and can be conditioned on age and gender. Evaluation on a test set shows that our model outperforms an optimized Gaussian process model in terms of predictive accuracy.