Modulation Spectrum Representations
Abstract
In this paper, we propose two different anomalous sound detection systems, one based on features extracted from a modulation spectral signal representation and the other based on i-vectors extracted from mel-band features. The first system uses a nearest neighbour graph to construct clusters which capture local variations in the training data. Anomalies are then identified based on their distance from the cluster centroids. The second system uses i-vectors extracted from mel-band spectra for training a Gaussian Mixture Model. Negative log-likelihood values are then used as anomaly scores. Both systems have been submitted to Task-2 of the DCASE 2020 Challenge and show significant improvement over the baseline AUC scores, with an average improvement of 6% across all machines. An ensemble of the two systems is shown to further improve the average performance by 11% over the baseline.