The problem of out-of-vocabulary words, more precisely proper names retrieval for in speech recognition is investigated. Speech recognition vocabulary is extended using diachronic documents. This article explores a new method based on neural network (NN), proposed recently by Mikolov. The NN uses high-quality continuous representation of words from large amounts of unstructured text data and predicts surrounding words of one input word. Different strategies of using the NN to take into account lexical context are proposed. Experimental results on broadcast speech recognition and comparison with previously proposed methods show an ability of NN representation to model semantic and lexical context of proper names.