Find Splices and Edits in Any Audio Recording
A free in-browser tool that hunts for splice points — places where audio has been cut, joined, or edited. Six forensic detectors (spectral flux, noise-floor change, ENF phase, pitch jump, timbre shift, sample-level clicks) vote on every 50 ms of audio and surface candidates with sample-precise timestamps.
How to detect splices
Upload your audio or video file (MP3, WAV, OGG, FLAC, M4A, AAC, MP4, WebM).
Wait for the six detectors to finish — typically 1–3 seconds for a few minutes of audio.
Read the verdict and heuristic score; clean recordings stay green.
Inspect the spectrogram timeline for red splice markers; click any marker to listen to the 1.5-second context.
Open "Why was this flagged?" on each splice card to see which detectors fired and why.
Export TXT or JSON if you need the report for a forensics workflow.
Find every cut and edit in any recording
How does this work?
Eight independent detectors run on a uniform 50 ms grid. Each one is normalised against the file's own statistics (μ + k·σ), then combined with reliability weights into a fused score. Peaks above the threshold are extracted as splice candidates and merged within 250 ms.
fused = Σ wᵢ · σᵢ(signalᵢ), splice_candidate = local_max(fused) ≥ 0.55
- Spectral flux discontinuity — STFT frame-to-frame change
- Noise-floor change-point — 10th-percentile RMS log-ratio
- Click — second-difference spike vs local RMS
- ENF (50 / 60 Hz) frequency + phase — quadrature detection plus parabolic-interpolated FFT peak; sub-Hz precision
- Pitch (F0) — autocorrelation, voiced-frame log-ratio
- MFCC — 26-mel, 13-coefficient L2 distance
- Bicoherence — normalised bispectrum, averaged over (f₁, f₂) ∈ [200 Hz, 2 kHz]
- RIR proxy — late-energy ratio (50–200 ms post-onset / peak) per 1 s window
When ENF or RIR is absent (outdoor recordings, dry mixes, continuous tones) their weight is redistributed proportionally across the remaining detectors so the fused score stays comparable across recording types.
References: Grigoras (2007) ENF criterion; Duxbury et al. (2003) spectral-flux onset detection; Davis & Mermelstein (1980) MFCC; Nikias & Petropulu (1993) higher-order spectra.
In-browser validation
Synthesises a battery of test scenarios in your browser and runs the detector on each, so you can verify our published metrics yourself. No test files are downloaded — everything is generated locally with Web Audio.
