Learning Temporal Dependencies in Connectionist Speech Recognition