Find out if a video clip or voice recording is real or AI-generated. TruthLens decodes visual keyframes with OpenCV, scores every frame through a neural vision ensemble, inspects temporal texture flicker & 2D FFT decoder grids, and runs 4-second sliding-window Wav2Vec2 audio deepfake detection.
Or click to browse local files from your phone or computer
Video: MP4, MOV, WebM, MKV • Audio: MP3, WAV, M4A, OGG, FLAC (up to 50 MB)
Almost no viral videos on X, TikTok, WhatsApp, or YouTube carry C2PA metadata. That is why TruthLens focuses directly on visual frame classification, temporal consistency, and acoustic vocoder forensics.
TruthLens decodes up to 8 evenly spaced keyframes across the video (plus a 70% center region-of-interest crop to ignore social media captions and black bars) and scores each frame through our vision classifier ensemble—displaying both the weighted score and the peak synthetic frame.
Generative video models (Sora, Kling, Runway Gen-3, Luma) and face-swap deepfakes struggle to maintain stable micro-texture sharpness across moving frames. TruthLens computes inter-frame Laplacian variance stability and compares central subject detail against the outer background.
When AI video generators upscale latent space frames into 720p or 1080p video, convolutional upsamplers often leave periodic 2D Fourier frequency spikes across keyframes. TruthLens inspects full-resolution keyframes for these mathematical grid anomalies.
Using ffmpeg, TruthLens extracts the video's audio track and evaluates it with a Wav2Vec2 deepfake voice detector across 4-second windows—catching cloned voices (ElevenLabs, XTTS, RVC) even when a scammer splices a few seconds of fake speech into real audio.
Each family of generative video and voice cloning models leaves distinct spatial, temporal, and spectral clues.
| Generator Family | Primary Visual or Acoustic Artifact | How TruthLens Detects It |
|---|---|---|
| Text-to-Video Diffusion (Sora, Kling, Veo, Runway Gen-3, Luma) | High-frequency texture swimming across consecutive frames, background object morphing, and synthetic latent diffusion smoothing | Multi-frame neural ensemble scoring + inter-frame Laplacian sharpness stability (CV) + 2D FFT spectrum |
| AI Talking Heads & Lip-Sync (HeyGen, D-ID, Wav2Lip, SyncLabs) | Over-smoothed lower-face/mouth box spliced onto a sharper static or looped head video | Center-subject vs. background Laplacian residual ratio + 70% center ROI frame classifier + audio track deepfake check |
| Face-Swap Deepfakes (DeepFaceLab, Roop, FaceFusion) | Frame-to-frame boundary flicker during rapid head turns or lighting shifts; mismatched skin grain | Peak synthetic keyframe detection across 8 sampled timestamps + 1-click keyframe OSINT reverse search |
| Neural Voice Clones & TTS (ElevenLabs, XTTS, Bark, RVC) | Neural vocoder phase artifacts, 16 kHz / 24 kHz spectral shelf cutoffs, or unnatural breath/pitch regularity | Wav2Vec2 4-second sliding-window classifier + Inferno STFT spectrogram + autocorrelation F0 pitch tracking |
Whether you are verifying a leaked political audio clip, a WhatsApp voice note, or an urgent voicemail, TruthLens visualizes the sound wave's internal physics alongside neural deepfake probabilities.
Sophisticated audio fakes often start with a real greeting before transitioning into a synthesized voice clone. By scoring overlapping 4-second windows independently, TruthLens pinpoints the exact timestamp window where synthetic speech peaks.
Many text-to-speech vocoders generate audio internally at 16 kHz or 24 kHz (capping frequencies at 8 kHz or 12 kHz) before saving into a 44.1 kHz MP3/WAV container. Our Inferno spectrogram highlights these brickwall frequency ceilings visually.
Autocorrelation pitch tracking measures the speaker's median vocal cord frequency (in Hz) and pitch deviation (±Hz), while Wiener entropy measures spectral flatness to distinguish tonal human harmonics from vocoder buzz.
TruthLens checks whether the video and audio content itself is real or AI-generated. Because 99.9% of videos online do not have C2PA metadata, we extract visual keyframes with OpenCV and decode the audio track with ffmpeg, running both through neural AI detectors, 2D Fourier frequency checks, and temporal consistency analysis. C2PA metadata is only checked as an optional secondary step.
When a video is uploaded to X, TikTok, WhatsApp, or Instagram, the platform re-encodes it using heavy H.264/HEVC block compression. Aggressive video compression blurs fine pixel textures—which can simultaneously hide subtle AI diffusion noise and introduce blocky compression artifacts on real camera footage. To combat this, TruthLens evaluates both full frames and 70% center crops, weights peak synthetic frames, and combines visual scores with audio deepfake detection.
Platforms like X, TikTok, and Instagram lock their video streams behind per-user session tokens and anti-bot rate limits, blocking third-party servers from downloading the raw MP4 file automatically. Simply save or download the video to your phone or computer and drop the MP4 file into the box above for full frame-by-frame and audio analysis.
Every video scan in TruthLens automatically extracts 6 visual keyframes across the clip's timeline and computes a 64-bit perceptual dHash for each frame. Click any keyframe in your report and use the 1-Click Keyframe OSINT Launchpad to search that exact frame on Google Lens and TinEye. This quickly reveals if an old video is being recirculated out of context or if it originated on an AI creator account.
Yes. Upload any .mp3, .wav, .m4a, .ogg, or .opus voice note directly into the scanner. TruthLens decodes the audio waveform, runs our Wav2Vec2 synthetic-speech classifier across 4-second windows, and generates a full STFT frequency spectrogram and vocal pitch report.
You can upload MP4, MOV, WebM, MKV, AVI, MP3, WAV, M4A, AAC, OGG, and FLAC files up to 50 MB. Audio tracks up to 180 seconds and evenly spaced video keyframes across the entire clip are analyzed in volatile memory and deleted immediately after your report is returned.
Learn how to spot visual AI giveaways in video frames, why social media strips hardware metadata, and how AI scripts sound.
Keyframe & Visual Checks
Metadata & Social Stripping
AI Script & Voice Cadence
Switch to a dedicated single-purpose workbench for still-image forensics, hidden prompt extraction, or document style analysis.
