| With the explosive amount of music data available on the internet in recent years, there has been much interest in developing new ways to search and retrieve such data effectively. Currently, most music search engines operate on text labels or symbolic data, rather than on the underlying acoustic content. A truly content-based music retrieval system should have the ability to find similar songs based on their underlying score or melody, regardless of their metadata description or file names. Potential applications include automatic music identification, music analysis, plagiarism detection, copyright enforcement, etc.; In this dissertation, we study the problem of searching and retrieving music based on acoustic content similarity. Given a query sound clip, our goal is to retrieve "similar" occurrences from a music database, where similarity is based on the intuitive notion of "same song" perceived by humans: two pieces are similar if they are fully or partially based on the same score, even if they are performed by different people, with different instruments, or at different tempo. Retrieval results are given as a list of songs ranked by computed similarity estimate. Both the input query and the underlying database are taken from actual music recordings in raw acoustic format.; We study two types of systems, one based on exhaustive matching by dynamic programming (which is relatively accurate but not scalable), the other based on high-dimensional indexing (which is less accurate but scalable). For the latter index-based retrieval system, the core algorithm is parallelizable and can be placed into a peer-to-peer architecture for improved performance, with the ability to share spare CPU resources and to achieve dynamic load-balancing. |
