Automated Sequence Clustering of Audio Signals using Conditional Random Fields

In this paper, we propose a new conditional random field(CRF) based algorithm for automated sequence clustering ofaudio signals, where the termsequence clusteringis usedin the same way as in biomedical signal clustering. UsualCRF based methods are trained in an observed manner, thatis, for the training data, we need a corresponding statesequence. For automated sequence clustering, we have nosuch known state sequence in the beginning. We thereforeadapt the training of the CRFs to the case of an unknownstate sequence and apply the trained model for classificationof new instances of the audio signals. This analysis of theacoustical signals can be used, for example, for scene analysisor novelty detection, where one uses abstract states regularlyand therefore manual prelabeling is not reasonable. In theexperiments, we successfully have applied the trained modelfor sequence clustering to audio signals and were able to detectthe significant clusters.