Water supply systems in Japan contribute significantly to improve public health. Unfortunately, there are many age-deteriorated pipes of various sizes and leaks frequently occur. Particularly devastating are hidden leaks occurring underground because when left undetected for years these leaks result in secondary damage. Thus, early detection and treatment of leaks is an important civil engineering challenge. At present the acoustic method is the most popular leak detection method. The purpose of this study is to propose an easy and stable leak detection method using the acoustic method assisted by pattern recognition techniques. In the proposed method we collect in the form of digital signals sound and pseudo-sound samples of underground leaking pipes. Principal component analysis (PCA) of the power spectrum of one leak sound is made, and a new coordinate system is constructed. We project the other sounds in the coordinate system, and evaluate if the sounds are similar to the sample sound or not by comparing the residual between the original and the projection. Next, we evaluate the DSF (Damage Sensitive Feature), which is a function of the first three AR model. At last, the feature vectors are created by combining the residuals, the DSF, and the damping ratio of the AR model, and a leak detection method is proposed using the Support Vector Machine (SVM) based upon them. In this study, it is shown that the residual and DSF are useful indices for leak detection. Furthermore, the proposed method shows high accuracy in recognizing leaks.