In this paper we present methods for identifying derivative
works in large audio collections, that is, works that contain
audio passages that resemble passages from a source work,
or set of source works. In our application, resemblance is
approximate, we do not look for exact matches of the signal.
This is because derivative works do not simply contain
``samples'' of the signal of an original work, but instead
use one aspect of the source, such as a vocal passage, and
remix it with new percussion and instrumental audio. Only
a very small part of the source work might be used for the
derivative work, so any method used to identify derivative
must identify sources in a completely new context, this is
called partial containment. Hence identification of derivative
works is partial containment of approximately matching
audio. Our solution uses audio shingling, a novel method
based on techniques for near-duplicate elimination in document 
search. We describe the steps for constructing and
comparing audio shingles, and for identifying partial containment 
using scalable hashing algorithm for nearest neighbour retrieval.