In a text-to-speech (TTS) system a proper prosody control is necessary for natural synthetic speech. But synthetic speech generated by regular rules can make a listener irritated and bored, because the implemented prosody is always same for the same sentence. In this paper, the randomness of the prosodic boundary index (PBI), as a primary study on irregularity of prosody is discussed. We examined the PBI data of 1,800 spoken sentences and concluded that there were irregularities. We applied the conventional stochastic method (CSM) to the PBI prediction. However, CSM could not implement PBI irregularity. So we proposed local constraint Viterbi search algorithm (LCVS) as an alternative method. LCVS was evaluated by computer simulation.