The paper presents algorithms and coefficients developed in order to select specific acoustic units to the base used in concatenative speech synthesis. The approach is based on the assumption that the database is created automatically. In the natural speech signal, which is a sample of the voice of a particular person, the acoustic unit must be marked and then cut out. This generates often very large, redundant collection of units from which the best units should be selected to the final base. Described coefficients refer to allophones databases in TTS synthesis.