Deepfake Audio Detection with Neural Networks Using Audio Features
Abu Qais, Akshar Rastogi, Akash Saxena, Arpit Rana, Deependra Sinha
Abstract
In this paper, a speech spoofing detection system based on Convolutional neural networks using different audio features has been proposed to classify the human speech and synthetic voice, Worst-case scenarios can develop using deepfake audios as threat to assets and image of a person, it can also become a threat to the whole country by unethical uses intended for loss of other party. Using a small voice clip of a person an attacker can develop similar voices. Every audio signal can be represented on a 2D graph plotted by mathematical calculations. The processing of audios into CNN requires a lot of computation, to make a system that can detect deepfake voices with much less computation by conversion of audios to images of audio features (Spectrogram, MFCC, FFT, STFT) and then obtaining the array values as a numeric format which are most suitable to feed. Different approaches for feeding data to model are applied for prediction individually as well as in a concatenated approach.