Standard correlation methods work poorly for significantly distorted signals. Good examples are 1-to 50-Hz underwater sounds correlated across hundreds of kilometers to locate ships, earthquakes, and large marine mammals. Current correlation methods for these sounds require either laborious human examination of graphical displays or automatic checking of enormous numbers of combinations. This paper reports on a program for automatically recognizing, classifying, and correlating such signals in a more efficient, higher level way. Ideas from computer vision are used to describe overall sound shapes rather than their details. The program uses a blackboard architecture and a time-shift transform analogous to the Hough transform. In experiments, the program ran an estimated 250 times faster than a conventional correlation approach. The program identified 69% of apparent whale moans having marked sound distortions with distance (and 100% of earthquakes), but nonetheless was nearly perfect in correlating whale moans between hydrophones 340 km apart in the presence of noisy earthquakes.