Listen, attend and spell: A neural network for large vocabulary conversational speech recognition | Synapse