Compacting Music Signatures for Efficient Music Retrieval
Authors
- Bin Cui (Peking University, China)
- H. V. Jagadish (University of Michigan, USA)
- Beng Chin Ooi (National University of Singapore, Singapore)
- Kian-Lee Tan (National University of Singapore, Singapore)
Abstract
Music information retrieval is becoming very important with the ever-increasing growth of music content in digital libraries, peer-to-peer systems and the internet. While it is easy to quantize music into a discrete string representation, retrieval by content requires (approximate) sub-string matching, which is hard.
In this paper, we present a novel system, called MUSIG, that uses compact MUsic SIGnatures for efficient content-based music retrieval. The signature is computed as follows: (a) each music file is split into a set of (overlapping) segments; (b) similar segments are clustered together; the number of clusters corresponds to the number of dimensions; (c) for each music file, the number of its segments that fall into a cluster determines the key value in that dimension.
Most index structures for multimedia are only able to provide an initial filtering and return a set of candidate answers that must be further examined. For MUSIG, we have also designed a scoring function that permits a ranked answer set to be generated directly based only on the signatures. Our experimental results show that this scheme retains a high degree of accuracy while being very efficient.