Abstract: This paper presents a novel double-talk detector (DTD) based on a nearest neighbor line (NNL) classifier. The underline idea is to use the feature information sufficiently and to design the DTD with pattern classification method. This paper analyzes 2 main kinds of conventional DTD, Geigel and DTD based on correlation, from the perspective of pattern classification, and then gives a new design method of DTD. A novel nonparametric classifier called NNL classifier is introduced to detect double-talk. NNL classifier has low computation cost and good performance. With NNL classifier, we fuse several conventional DTD and avoid the problem of making a fixed threshold, which exists in most of the conventional DTD. So the NNL-DTD is robust in adverse conditions. Experiments show that the proposed approach is more effective than conventional methods.