Topic detection is the process of finding the topics in a document collection. For a large amount of dataset, manual topic detection is difficult or even impossible. Thus, we need an automatic method known as Topic Detection and Tracking (TDT). One of the TDT methods used for topic detection problem is a clustering-based method such as fuzzy C-means (FCM). FCM works reasonably well on low-dimensional data but fails on high-dimensional data. In the high-dimensional data, a random-based initialization of FCM converges to one cluster center called center of gravity, so that all topics generated are similar. In this paper, we examine a non-random initialization by using singular value decomposition (SVD). Our simulations show that the SVD-based initialization method solves the center of gravity problem in a certain degree of fuzziness and gives a better accuracy than the random-based initialization.