为了提高两个说话人情况下电话语音双人检测的性能,提出了一种基于长时距离与短时距离之和的双人语音分段算法,并对分段过程中的聚类过程进行了改进。对于分段得到的语音采用基于GMM-UBM(Gaussian mixturemodel-universal background model,GMM-UBM)模型的说话人确认算法,构成一个电话语音双人检测系统。实验结果表明:在美国国家标准技术局1999年组织的年度说话人识别评测双人检测数据库上的检测等错误率为15.1%。该算法取得了良好的效果。 In order to improve the performance of double detection of telephone voice in two speaker scenarios, a two-voice segmentation algorithm based on the sum of long-distance and short-distance is proposed and the clustering process in segmenting is improved. For the segmented speech, a speaker recognition algorithm based on the GMM-UBM (GMM-UBM) model is adopted to form a telephone voice double detection system. The experimental results show that the error rate of detection and so on on the double detection database of the annual speaker recognition evaluation organized by the United States National Bureau of Standards and Technology in 1999 is 15.1%. The algorithm has achieved good results.