Semantic Veracity Analyzer: Detecting Vocal Inconsistencies via FFT Peak Gradient Analysis for ASR Pre-filtering, AI Self-Feedback, and Psychiatric Voice Biomarker Applications

Open-source Python system that detects vocal inconsistencies through Fast Fourier Transform (FFT) spectral decomposition and multi-feature gradient analysis. The system combines peak amplitude gradient consistency, MFCC delta variance, spectral flux, Lippold microtremor energy (8-12 Hz), and Praat-derived vocal features (jitter, shimmer, HNR, formants F1-F4) into a Naturalness × Involuntary Stress Cartesian mapping.

Key finding: deceptive speech exhibits reduced voluntary variation (jitter -27%, MFCC delta variance -40%, spectral flux -52%) but elevated involuntary microtremor (+17%) — consistent with the over-control hypothesis of deception (Zuckerman et al., 1981).

Three proposed applications: (1) ASR truth pre-filtering to reduce hallucinations from contaminated input, (2) AI self-feedback where TTS systems analyze their own synthetic voice for uncertainty signals, (3) psychiatric voice biomarker monitoring for mood disorders.

Includes a section on synthetic voice signal equivalence — demonstrating that FFT processes synthetic and organic voice identically at the signal level — and the proposal of "Logos Probabilis" as a taxonomic framework for probabilistic intelligence systems.

Co-authored with Claude (Anthropic), who contributed Section 6 (AI perspective on self-feedback) and Section 7 (signal equivalence hypothesis).

Code: https://github.com/sfaustodev/NLP-AI