AFIP's Voice and Audio Forensics division develops methodology for detecting AI-cloned voices, synthetic speech, and manipulated audio recordings. As voice synthesis technology advances, forensic detection requires increasingly sophisticated spectral and biomarker analysis.
Voice cloning systems replicate the perceptual qualities of a target voice but introduce artifacts in the spectral domain. AFIP's spectral analysis framework examines formant distributions, harmonic-to-noise ratios, and micro-temporal variations that distinguish biological vocal tract output from synthetic generation.
Our analysis operates across multiple frequency bands simultaneously, detecting inconsistencies that single-band analysis would miss.
Natural human speech contains biomarkers — micro-variations in pitch, timing, and spectral energy reflecting physiological processes. Current voice cloning replicates macro-level characteristics but fails to reproduce the full complexity of biological signals.
Our research has identified 14 distinct biomarker categories providing independent detection signals. Combined biomarker analysis achieves detection rates exceeding 96% across tested platforms.
AFIP maintains an ongoing benchmark program testing detection accuracy against commercial and open-source voice cloning systems. Current results: 96.2% detection accuracy with a 1.4% false positive rate across 12,000+ analyzed samples.