AUDIONYX: REAL-TIME DETECTION OF AUDIO DEEPFAKES IN PHONE CALLS

Abstract

The explosion of AI-assisted voice synthesis technologies has made audio deepfake–based fraud a greater risk, especially within telecommunication domains. These synthetic voices are one of the leading impersonation methods, attacks and scams with potentially grave security hazards. Detecting real-time deepfakes is challenging due to bandwidth limitations, codec compression, and background noise that obscure distinguishing artifacts. This paper presents Audionyx, a real-time deepfake detection framework for telephony applications. It uses a lightweight custom Convolutional Neural Network (CNN) trained on Melspectrogram abstractions to strike an optimal balance between accuracy in detection and computational efficiency. A sliding window segmentation strategy and probabilistic aggregation mechanism ensure stable and reliable detection across continuous audio streams. Experimental evaluation demonstrates excellent detection performance and low latency, testing the ability of the system to be deployed in real time. The proposed approach is a robust and scalable method for reducing fraud through voice and for improving security against impersonation attacks.