Speech segmentation is the foundation of some applications such as speech recognition and spoken document retrieval. An improved algorithm is proposed here which include: (1) GLR variance based threshold adaptive algorithm is to improve the threshold selection approach in speaker based speech segmentation under various acoustic environments;(2) BICs Detection Ability is referred to determine when BIC is effective;(3) Besides to verify the candidate segmentation points, BIC is used to calibrate their bias caused by GLR variance.Experimental results indicate that the improved algorithm is prior to the original one.