We propose a new method for speech synthesis that concatenates waveforms selected from a waveform dictionary. The method uses a modified PSOLA technique to alter the pitch of waveforms selected from a dictionary. The limits of acceptable pitch shifts are determined by preference tests. To make segment selection more accurate, we introduce a new factor which considers the spectral continuity across voiced phoneme boundaries. The average spectral difference is reduced from 5.4dB to 2.7dB and the synthesized voice is more fluent.