In the paper we present our experiments with automatic pruning of speech databases created by us for Unit Selection based speech synthesis systems. Several algorithms have been attempted and perceptually evaluated. An optimal size of speech database has been reached where lose of naturalness due to unit pruning is not perceptible.