Over the past decade, many research fields have real- ized the benefits of motion capture data, leading to an ex- ponential growth of the size of motion databases. Conse- quently indexing, querying, and retrieving motion capture data have become important considerations in the usability of such databases. Our aim is to efficiently retrieve mo- tion from such databases in order to produce real-time an- imation. For that purpose, we propose a new database ar- chitecture which structures both the semantic and raw data contained in motion data. The performance of the overall architecture is evaluated by measuring the efficiency of the motion retrieval process, in terms of the mean time access to the data.