Abstract. The present paper deals with choosing the base type for the unit selection speech synthesis method of the Lithuanian language. Phoneme and diphone units have been examined. Besides, two different methods of joining costs calculation were employed in a diphone synthesizer: one was based on the spectral similarity and the other was based on phonological classes of the sounds to be joined. Synthesizers were evaluated according to their performance, algorithm complexity, the number of joins in a synthesized speech and the human listeners' subjective judgment. Experimental testing showed that the diphone synthesizer based on phonological classes was much more acceptable to the listeners than the one based on the spectral similarity. The diphone synthesizer based on phonological classes outperformed the phonemic synthesizer in terms of performance and the number of joins though it was somewhat less acceptable to human listeners.