Research, Insights & Forensic Science
In-depth technical papers, low-difficulty practical guides, and linguistic benchmarks examining generative AI models, vocoder acoustics, and academic integrity.
Acoustic Forensic Guide: How to Tell if Audio is AI Generated Using Spectral Cutoffs & Jitter
Generative voice cloning systems like ElevenLabs, GPT-4o Advanced Voice, and Suno leave microscopic mathematical footprints. We unpack how 16 kHz brickwall frequency filters, pitch jitter flatlines (0.012%), and missing glottal pulse variations expose deepfake synthetic audio under laboratory Mel-spectrogram analysis.
Understanding Neural Perplexity: Why Human Writing Outperforms AI on Readability & Flow
How modern university classifiers (Turnitin v3, GPTZero) compute sentence perplexity and burstiness. Learn why mechanical synonym swapping fails while genuine structural variation mirrors biological human prose.
The Death of Monotone Pitch: Analyzing ElevenLabs v3 and OpenAI Advanced Voice
While first-generation text-to-speech sounded robotic, latent diffusion audio models simulate emotional inflection. Our lab demonstrates how phase-shift anomalies expose their digital origins.
High-Frequency Wavelet Analysis: Catching Diffusion Latents in Photorealistic Portraits
A microscopic dive into Fourier transforms and Discrete Cosine Transform (DCT) patterns that reveal neural latent upscaling in Flux, Midjourney v6, and Stable Diffusion 3.5.
Zero-Retention Auditing: Why Students Need Local Ephemeral Scanning
Traditional checkers upload student dissertations to central databases, inadvertently creating intellectual property traps and future false plagiarism alerts. Here is how ephemeral RAM verification resolves it.
The False Positive Dilemma: Establishing Court-Grade Standards for AI Detection
Why probabilistic percentages alone cannot constitute disciplinary evidence in universities or legal proceedings. A comprehensive framework for biometric, semantic, and cryptographic validation.
C2PA vs. SynthID: The Battle for Cryptographic Provenance in Generative Media
Comparing open coalition cryptographic manifests (C2PA) against imperceptible latent watermarks (Google SynthID). Can metadata survive social media transcoding and audio re-sampling?