Audio Detection Concepts

AI-generated audio (deepfakes) has improved significantly, but often lacks the natural imperfections of human speech.

Spectral Analysis

SpectralSignal analyzes the frequency spectrum for artifacts. Synthetic audio might lack high-frequency details or have unnatural spectral continuity.

Physiological Signs (Breathing)

BreathingSignal looks for natural breathing pauses. Deepfake models often generate continuous speech without the natural pauses humans need to breathe, or they insert them in unnatural places.

Foundation Models (Wav2Vec, AASIST)

Using large pre-trained models like Wav2Vec or specialized architectures like AASIST (for anti-spoofing) allows detecting subtle patterns learned from massive datasets of real and spoofed audio.