The ability of the ear model and lateral inhibitory networks (LIN) to preserve and enhance acoustic features of speech signal is examined by training neural networks to recognize phonemes by their LIN outputs. Using the back propagation learning algorithm, networks that are specialized to recognize specific classes of phonemes are trained and tested. Experiments are conducted both in single and multi-speaker cases. By using single layer networks, we can show that the phonemes are identified by their acoustic features that have been known to linguists and phoneticians. The networks generally yield satisfying results when tested in experiments for a single speaker, where we focus on the performance against phoneme variation induced by the context, and in multi-speaker experiments where errors in recognition are due to speaker variation. These results convince us that the acoustic features picked by the networks are reliable cues for phoneme recognition