In order to develop a speech translation system that has a large vocabulary and accepts various expressions, we have carried out a practical and quantitative investigation of similarities and differences among the tasks of machine translation for written language such as newspaper articles, and those for spoken language such as conversations in daily life. First, we mention the characteristics of vocabularies for colloquial expressions in conversations. Next, we report the characteristics of basic sentence patterns which consist of one predicate and essential case phrases from the viewpoint of translation from Japanese to English. Finally, we discuss state-of-the-art technology based on a case study of a Japanese-to-English speech translation system ATR-MATRIX built by ATR Interpreting Telecommunications Research Laboratories.