Our core detection engine is powered by our lightweight AASIST-X spectro-temporal graph attention architecture (`759K` parameters) paired with custom 1D LIME temporal attribution. Because the full research manuscript is currently reserved prior to formal peer-reviewed publication, full mathematical derivations, algorithm pseudocode, and complete benchmark tables are withheld from public download on this page.
4-Stage Live Verification Pipeline
Built for Real-Time Results16 kHz Raw Waveform Capture
Standardizes live microphone streams, browser audio, and uploaded files into 16,000 Hz 16-bit mono PCM without lossy spectrogram compression.
Competitive Feature Gating
Suppresses background room noise and channel redundancy while preserving fine-grained phase transitions across residual blocks.
Spectro-Temporal Attention
Weights both spectral vocal-tract formants and temporal cadence simultaneously to isolate fleeting vocoder or voice-conversion glitches.
1D LIME Time Attribution
Partitions the recording into phonetic time windows and highlights exact authentic (emerald) vs. synthetic (crimson) segments.
1D LIME Temporal Attribution Map
Onset breath & glottal attack
Formant transition F1/F2 stability
High-frequency vocoder phase check
Pitch-synchronous micro-jitter
Sibilant fricative energy (4–7 kHz)
Inter-syllabic silence floor
Harmonic-to-noise continuity
Trailing decay & room reflection