Bi-Directional Recurrent Neural Networks For Speech Recognition

While many possible network architectures have beenused to estimate conditional probabilities of classmembership, recurrent neural networks (RNNs) havebeen most successful for speech recognition. In thepast optimal results were achieved by merging theoutputs of two RNNs trained in each time direction.Merging outputs of different experts to form one resultingopinion is theoretically difficult - a direct combinationof the experts during training would be desirable.This paper presents a...