We studied fully recurrent networks trained to recognize temporal patterns that resembled sequences of pure tones. It is argued that the notion of a temporal pattern recognition implies a method for measuring time, of which there appear to be 3 standard methods: sequence or serial order, absolute measurement in some clock units, and phase angle measurement (as fractions of a period present in the stimulus). The 'representations' that were developed by the networks took the form of trajectories through activiation state space that were sculpted by a chain of stable attractors. The training procedure (real-time recurrent learning) moved the locations of the attractors for pattern elements and the location of the 'recognition region' (the hyperplane where an output node has the value 1) in such a way that sufficient information about the history of the inputs is retained to differentiate targets from distractors. We demonstrate some capabilities and limitations of these models for an understanding of human hearing.