Speech recognition with neural networks and network fusion

Large neural networks can be segmented into several small subnetworks, trained independently, and combined. In the present work, a rule termed network fusion is derived for combining these subnetworks. This rule can be implemented with a single-layer neural network and is optimal if the output of the subnetworks is independent. The rule is demonstrated on a Gaussian random process and applied to a small vocabulary speech recognition system. The proposed method of combining the separately designed subnetworks into a large network exhibits a performance that approaches the Bayes limit