Research fields such as speech recognition require a large amount of speech data with phoneme label information uttered by various speakers.Ho wever, phoneme labeling by visual inspection segmentation of input speech data into corresponding parts of given phoneme by human inspection is a time-consuming job.An automatic phoneme labeling system is required.Cur rently, several automatic phoneme labeling system based on Hidden Markov Model(HMM) were proposed. The performance of these systems depends on the used phoneme models.In this paper, at first, we propose an acquisition algorithm of accurate phoneme model with the optimum architecture, and then the obtained phoneme models is applied to segment an input speech without phoneme label information into the part corresponding to each phoneme label