Building HMM based unit-selection speech synthesis system using synthetic speech naturalness evaluation score