In this paper, a parametric trajectory segment model (PTSM) with segmental inner time warping is proposed to improve the recognition accuracy of large vocabulary continuous speech recognition(LVCSR). The proposed PTSM utilizes the state boundary information provided by HMM system during decoding to do segmental inner time warping. Good alignment between different length realizations of a same phone unit can be obtained by this method. Based on the effective alignment, a new distance measure of measuring the average value of the norm of the residual error is used in k-means clustering to decide the parameters of the mixture density of PTSM. For two LVCSR tasks, the HMM system working with the proposed PTSM can give a consistent improvement over either the HMM system working with the traditional PTSM or the HMM system working alone.