How similar do songs sound? Towards modeling human perception of musical similarity

Human listeners have a well-developed feeling for identifying "whether two songs sound similar" or whether they do not. Even though this type of judgment usually also involves a considerable amount of the listener's background knowledge, it has been demonstrated that an algorithmic model of this type of similarity can be achieved by merely evaluating the signal's low-level acoustic features. The paper describes a system for assessing subjectively sound similarity between pairs of musical items by using a number of such signal features. The system's performance is assessed by means of a subjective listening test that is based on a modification of a test methodology originally standardized for subjective sound quality evaluation. A number of interesting applications for such a technology are described.