Abstract:
When detecting anomalies in audio, it can often
be necessary to consider concept drift: the distribution of the
data may drift over time because of dynamically changing
environments, and anomalies may become normal as time
elapses. We propose to use dynamic Huffman coding for
anomaly detection in audio with concept drift. Compared with
the existing method of adaptive Gaussian mixture modeling
(AGMM), dynamic Huffman coding does not require a priori
information about the clusters and can adjust the number of
clusters dynamically depending on the amount of variation in
the audio. To control the size of the Huffman tree, we propose
to merge clusters that are close to each other instead of
replacing rare clusters with new data. This reduces redundancy in the Huffman tree while ensuring that it never forgets
past information. On audio datasets with concept drift which
we have curated ourselves, our proposed method achieves a
higher area under the curve (AUC) compared with AGMM and
fixed-length Huffman trees.
