A selective use of non-uniform synthesis units for speech synthesis-by-rule is discussed focusing on an optimal unit selection method. In this paper, we propose two algorithms for unit selection. The first one uses one total measure reflecting contextual similarities and adequacy of unit concatenation. The second one combines top down control for concatenation points and bottom up search for the appropriate speech template. The high quality of both selection methods, compared to the conventional method using fixed units, is confirmed by both subjective and objective tests. Furthermore, the results of intelligibility tests are analyzed aiming at designing a quantitative measure to evaluate unit suitability.