One issue arising in text-to-phone conversion is inconsistency between its output and the phonetic time-alignment of the dataset, hindering the back-end's ability to access the best units to synthesize a text. Some such inconsistency is inevitable because dataset labeling requires allowance for alternate pronunciations of words, while the front-end typically predicts a single pronunciation for a word. In this paper we describe an alternate algorithm that recovers from these inconsistencies. The front-end is modified in order to allow multiple pronunciations for a word. The selection of the best pronunciation is based on an a posteriori cost function comparison between the synthetic signals.