This paper presents the design and early recording stages of a Romanian speech database to be used for development of both speech recognition and speech synthesis systems. The recognition part is built around a core patterned after the EUROM_1 [4] design, so that an as good as possible compatibility to exist with this, and includes both read and semispontaneous speech. The synthesis part consists of a read speech corpus from which diphones are to be extracted to build concatenation-based TTS systems, and read material to serve as benchmark data for the administration of a Romanian version of the Modified Rhyme Test [2].