The Use of Regression Models for Detecting Digital Fingerprints in Synthetic Audio
Abstract
Modern advancements in text to speech and voice conversion techniques make it increasingly difficult to distinguish an authentic voice from a synthetically generated voice. These techniques, though complex, are relatively easy to use, even for non-technical users. It is important to develop mechanisms for detecting false content that easily scale to the size of the monitoring requirement. Current approaches for detecting spoofed audio are difficult to scale because of their processing requirements. Individually analyzing spectrograms for aberrations at higher frequencies relies too much on independent verification and is more resource intensive. Our method addresses the resource consideration by only looking at the residual differences between an audio files smoothed signal and its actual signal. We conjecture that natural audio has greater variance than spoofed audio because spoofed audios generation is conditioned on trying to mimic an existing pattern. To test this, we develop a classifier that distinguishes between spoofed and real audio by analyzing the differences in residual patterns between audio files.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2022
- Accession Number
- AD1184800
Entities
People
- William B. Brown
Organizations
- Naval Postgraduate School