Short utterance recognition using a network with minimum training