This paper addresses the problem of query-by-example spoken term detection (QbE-STD) in the presence of background noises that are inevitable in real applications. To deal with this, we propose a convolutional neural network (CNN) based bottleneck feature representation for a keyword. A combined network that is made by attaching a bottleneck layer on top of a CNN is trained on Wall Street Journal (WSJ) database. Finally, dynamic time warping (DTW) based template matching is performed to measure the distance between enrollment and test feature matrices which are extracted from the bottleneck layer. The proposed method is evaluated in terms of equal error rate (EER) on Aurora 4 Database. A series of experimental results verify that the proposed method performs significantly better than the baseline system in noisy environments shows over 30% relative equal error rate (EER) improvement in average.