This paper proposes a new speech recognition scheme for foreign out-of-vocabulary words embedded in native-language speech. To recognize foreign names frequently observed in news speech or in translation speech, we adopted a hierarchical language model that had been successfully applied to OOV words covering native vocabularies. In this hierarchical language model, OOV vocabularies are modeled as a word-class model in the upper-layered model, and their statistical phonotactic constraints are modeled in the lower-layered model. Since extra statistics are needed to cover foreign words and their pronunciation differences, we have introduced two techniques. The first is to combine translation target language models and translation source statistics of OOVs using the hierarchical language model. The second is to automatically generate recognition target pronunciations from original pronunciations by syllable-to-syllable mapping. To confirm the validity of this recognition scheme, we have conducted speech recognition experiments using English speech including Japanese personal names as OOV words. The proposed method outperformed the existing algorithm using a lexicon consisting of all the words in the training set. Surprisingly, it achieved better OOV recognition results than the non-OOV condition where all the proper names in the test set are registered in the lexicon.