Deepfake Speech Detection: A Spectrogram Analysis

Abstract

The current voice biometric systems have no natural mechanics to defend against deepfake spoofing attacks. Thus, supporting these systems with a deepfake detection solution is necessary. One of the latest approaches to deepfake speech detection is representing speech as a spectrogram and using it as an input for a deep neural network. This work thus analyzes the feasibility of different spectrograms for deepfake speech detection. We compare types of them regarding their performance, hardware requirements, and speed. We show the majority of the spectrograms are feasible for deepfake detection. However, there is no general, correct answer to selecting the best spectrogram. As we demonstrate, different spectrograms are suitable for different needs.The current voice biometric systems have no natural mechanics to defend against deepfake spoofing attacks. Thus, supporting these systems with a deepfake detection solution is necessary. One of the latest approaches to deepfake speech detection is representing speech as a spectrogram and using it as an input for a deep neural network. This work thus analyzes the feasibility of different spectrograms for deepfake speech detection. We compare types of them regarding their performance, hardware requirements, and speed. We show the majority of the spectrograms are feasible for deepfake detection. However, there is no general, correct answer to selecting the best spectrogram. As we demonstrate, different spectrograms are suitable for different needs.