Contents
Copy-move forgery in audio, also called cloning, duplicates a segment of a recording and pastes it elsewhere in the same file, to repeat a word, pad a silence, extend applause, or change what was said. Detection works by finding two stretches that are suspiciously identical, and audio has a sharper test than images do: if two intervals cancel to near-silence when one is subtracted from the other, one was copied from the other. Yan, Yang and Huang give the canonical example, where “the sentence ‘I’m guilty’ can be changed to ‘I’m not guilty’” by copying and inserting the word “not” spoken by the same voice elsewhere in the recording (IEEE TIFS, 2019). This is the audio counterpart to clone and copy-move forgery detection in images.
Why cloning needs its own test
A copy-move edit defeats the checks built to catch a splice. A splice joins two different recordings, and the seam between them shows up as a break in the device, room, mains-hum or codec traces that should stay constant across one continuous capture. Copy-move has no such seam, because the pasted material comes from the same file: it carries the same microphone coloration, the same background, the same electrical hum and the same compression history as its surroundings, so every continuity check stays consistent and misses it. The one thing a clone cannot hide is that it is an exact repeat of audio that already exists in the file, which is what a dedicated duplication test looks for.
The categorical result: signal subtraction
That is why audio copy-move is the closest an examiner gets to a demonstrable finding. The ENFSI audio manual treats a time interval “in which the content is perfectly identical” to another as “a strong sign of human post-processing,” and it prescribes the check: “The presence of a replicated portion may be verified using signal subtraction: if the signal is cancelled out by this process a copy/paste might have” occurred (ENFSI, 2022). Natural speech does not produce two perfectly identical waveforms at two different moments, because breath, mouth position, room response and background noise all differ even when a person says the same word twice. Exact cancellation therefore has no innocent production explanation in evidentiary speech, unlike a repeated musical loop, which makes the subtraction check deterministic in a way that no image tamper test quite matches. The same manual requires that any automated hit be confirmed by ear before it is trusted.
The detectors that find the pairs
Several automated detectors find the candidate pairs before the subtraction step, and real edits rarely leave a perfect copy to subtract, so they search for near-duplicates. Imran and colleagues (2017), noting that “blind detection of such tampering in digital audio is mostly unexplored,” extracted words with a voice-activity detector and matched them by one-dimensional local-binary-pattern histograms, reaching 96.59 percent and holding that accuracy after noise was added. Maksimović, Aichroth and Cuccovillo (2021) matched audio fingerprints without pre-segmentation, reporting precision and recall of 0.978 and 0.981 on clean material at a localization tolerance of 0.15 seconds, and still 0.935 and 0.942 when the replica was corrupted with white noise at 30 decibels. Yang, Liu and Cao (2023) reformulated the search to work fully blind and fast, localizing duplicates “even those as short as a fractional second” at about 99.70 percent average precision and a real-time factor of 0.008, quick enough to triage long recordings. Yan, Yang and Huang (2019) targeted robustness directly, matching the pitch and first two formant tracks of voiced segments by dynamic time warping, on the finding that “pitch and formant can be used as the features representing a voiced speech segment, and these two features are very robust against commonly used post-processing operations.” At their chosen threshold, 99.47 percent of genuine duplicate pairs fell below it while only 0.81 percent of non-forged pairs did, and clean, the method reached 99.48 percent precision at 100 percent recall.
What is different from image copy-move
The contrast with the image case is the interesting part. Image copy-move detection relies on matching features, either dense blocks of transform coefficients or sparse keypoints like SIFT, and a forger defeats it with geometric edits, scaling, rotating or flipping the cloned region so the features no longer match. Audio’s equivalent scrub is to shift the pitch of the copied segment, which moves the very pitch and formant tracks the robust detectors compare. It is the strongest published copy-move attack, and Yan and colleagues confirm it “leads to an obvious performance degradation,” but it degrades rather than defeats the method: pitch-shifting drops their precision and recall to roughly 78.52 and 75.28 percent for a raised pitch and 81.07 and 77.42 percent for a lowered one, not to zero, and a shift large enough to break the match leaves audible artifacts of its own. Where a geometric edit can quietly defeat an image clone matcher, an exact audio duplicate stays provable by subtraction, and disguising it costs the forger audible quality.
Where false positives come from
The natural false positive is genuine repetition. A word read twice, a repeated phrase, a looped sample or a stretch of applause can be legitimately similar, which is the audio counterpart to the repeated brickwork and foliage that fool image clone detectors, and it is exactly why auditory confirmation is mandatory rather than optional. The interpretation depends on content: a repeated chorus in a song can be normal, while a repeated word in a witness recording is not. The method also has blind spots, notably the duplication of unvoiced sounds, which carry no pitch to match, and very short duplicates below a detector’s granularity, though the fast blind methods push that floor under a second.
A signal, not a verdict
Copy-move detection answers one narrow question well, has a stretch of this recording been duplicated inside it, and on clean audio it answers it about as firmly as audio forensics gets. On a clip that has been heavily re-encoded off a messaging app, even a clean-looking result can mean only that too little signal survived to test. Forensic reporting requires any such finding to be stated as a strength of support for one proposition over another, never as proof (ENFSI, 2015), and confirmed by ear before it is trusted. It is one of the edit traces an examiner works through in what forensics can learn from a recording.
Sources
- Imran, Ali, Bakhsh, Akram (2017). Blind Detection of Copy-Move Forgery in Digital Audio Forensics. IEEE Access 5:12843-12855. DOI: 10.1109/ACCESS.2017.2717842
- Maksimović, Aichroth, Cuccovillo (2021). Detection and Localization of Partial Audio Matches in Various Application Scenarios. Multimedia Tools and Applications 80:22619-22641. DOI: 10.1007/s11042-020-09912-4
- Yang, Liu, Cao (2023). Fast and Blind Speech Copy-Move Detection and Localization in Noise. arXiv:2302.07584.
- Yan, Yang, Huang (2019). Robust Copy-Move Detection of Speech Recording Using Similarities of Pitch and Formant. IEEE Transactions on Information Forensics and Security. DOI: 10.1109/TIFS.2019.2895965
- European Network of Forensic Science Institutes (2022). Best Practice Manual for Digital Audio Authenticity Analysis, ENFSI-FSA-BPM-002.
- European Network of Forensic Science Institutes (2015). ENFSI Guideline for Evaluative Reporting in Forensic Science (STEOFRAE).