Even though the ideal split policy for M-tree should determine which two routing objects are promoted so that the two partitioned regions would have minimum volume and minimum overlapped space, the previously proposed split algorithms must fail to achieve the goal since they select the actual entries from overflow entries as the routing objects to be promoted. Hence, in this paper, we propose an optimal node split algorithm for M-tree which uses the virtual center points of two partitioned regions as the routing objects and which results in tighter volume and finally reduce the overlap index spaces between nodes. Also, we show that the proposed algorithm outperforms the previous ones for similarity-based query processing with respect to the average response time, although there is a slight increase in insertion time. We also discuss the implementation of a prototype of contentbased music retrieval system in which the proposed scheme is used for multidimensional indexing.