A neural network approach to predicting the reverberation time, RT60, at the conceptual design stage of auditoria, and churches is presented. The results of investigations previously carried out indicated that there was a good basis for using trained neural networks to predict the reverberation time for unoccupied enclosures but that 15 input variables were required to achieve the desired accuracy. As the number of input variables that can be readily identified and quantified at the early design stage is small, the objective of this work is to reduce network size and to obtain optimal neural networks. The results showed that the generalization performance of neural networks with simplified internal representation is efficient. Generally, the reverberation time prediction accuracy of the network models, for the six enclosures ‘tested’, is within the range of the subjective difference limen (ΔT/T ≈ 5%).