Clustering Algorithms Study for Audio Insertion Detection

In this paper, we reflect on a problem of audio editing detection. We consider a certain type of editing - the insertion of falsified fragments into an audio recording. To detect such insertions, we propose an algorithm based on audio background noise clustering. For the proposed algorithm, various methods for background noise preprocessing, various neural feature extractors and various clustering algorithms are examined. Based on the experimental research, we select a combination of a preprocessing method, a feature extraction model and a clustering algorithm that gives the best clustering metrics. The combination of ResNet feature extractor and agglomerative clustering yielded the best results.