“What are You Listening to?” Explaining Predictions of Deep Machine Listening Systems

Researchers have proposed methods to explain neural network predictions by building explanations either in terms of input components (e.g., pixels in an image) or in terms of input regions (e.g., the area containing the face of a Labrador). Such methods aim to determine the trustworthiness of a model, as well as to guide its improvement. In this paper, we argue that explanations in terms of input regions are useful for analysing machine listening systems. We introduce a novel method based on feature inversion to identify a region in an input time-frequency representation that is most influential to a prediction. We demonstrate it for a state-of-the-art singing voice detection model. We evaluate the quality of the generated explanations on two public benchmark datasets. The results demonstrate that the presented method often identifies a region of an input instance that has a decisive effect on the classification.