In this work, we are interested in assessing the optimality of the human auditory system, when the input stimuli is natural speech that is affected by additive noise. In order to do this, we consider the DANTALE II listening test paradigm of Wagener et al., which has been used to evaluate the intelligibility of noisy speech by exposing human listeners to a selection of constructed noisy sentences. Inspired by this test, we propose a simple model for the communication and classification of noisy speech that takes place in the test. We then identify a number of key properties that the test subjects satisfy, and combine these with our proposed model in order to derive optimal classifiers in the sense of maximum a posteriori estimation. We finally compare the performance of the classifiers to that of humans on the same noisy test sentences. The results reveal that at low SNRs, the human performance is inferior to that of the optimal classifiers. We conclude that in this special task, the human auditory system is not optimal.