Audio Source Verification Method Based on Structural Re-parameterization Network

Abstract

With the development of science and technology, the application of audio data in law enforcement and judicial fields is becoming increasingly widespread. Multimedia data such as audio recordings are prone to forgery, which brings great trouble to judicial fairness. When digital recordings are used as evidence, effective techniques are needed to ensure their reliability. For example, digital audio needs to verify its recording device. In this paper, we focus on the verification problem of audio source, i.e. determining whether a piece of audio recording comes from a given target device. We propose an audio source detection framework based on a structural re-parameterized network, and with a carefully designated loss function, the recognition accuracy is improved under the noise conditions. Experiments show the proposed method achieved a TPR of 99.89% and an FPR of 4.17%, which is superior to existing audio source detection methods.