How can AI-generated voices be identified and combatted to prevent disinformation?

In this study, methods for identifying AI-generated voices to combat the spread of disinformation were investigated. The methods used were based on a qualitative case study, employing technical analyses of three public cases where AI-generated voices were used to spread disinformation. These recordings were compared with authentic voice recordings of the same individuals, utilizing waveform analysis and spectral analysis. Spectrograms were studied visually and complemented by quantitative acoustic measures such as High-Band Energy Ratio (HBER), spectral flatness, spectral centroid standard deviation, and Signal-to-Noise Ratio (SNR). The analysis showed that AI-generated voices generally exhibit lower energy in high-frequency regions and less variation in spectral characteristics compared to authentic speech. However, results varied slightly between the cases, suggesting that the technology and quality of AI-generated voices depend on the specific methods and models used. Although some acoustic measures proved effective in identifying AI-generated voices, the study indicated that individual measures alone are not always sufficient for reliable identification. Rather, a combination of various measures and analytical methods is necessary. The study concludes that technical analysis based solely on spectrograms and a limited set of acoustic measures cannot reliably identify AI-generated voices in all cases. Consequently, as the technology evolves, society must continuously develop both technical and legal methods to effectively counteract the spread of disinformation through AI-generated voices.