Contents
An edit to a recording leaves seams, and audio tampering detection is the set of techniques that find them. The two that matter most are the splice, a cut or insertion that breaks a trace which should run continuously through one capture, and the copy-move, a segment duplicated inside the same file to repeat, pad or alter what was said. Neither is ever reported as proof, because innocent events can trip the same alarms, and the governing standard states the limit plainly: “evidence of discontinuities is not evidence of editing” (ENFSI). What follows is what an examiner looks for, and how far each read holds.
What counts as tampering
Tampering means the content was changed after capture: a word removed, a phrase inserted, two recordings joined, a passage duplicated. It is distinct from ordinary processing such as normalization or format conversion, which alters the whole file uniformly and leaves no internal seam. Detection therefore hunts for local inconsistency, a point where one stretch of the recording does not belong with the rest, and the examiner’s task is to decide whether that inconsistency is better explained by an edit or by something innocent that happened during capture.
How is a splice caught?
A butt-splice or insertion is found three ways, and their strengths trade off against compression: the sharpest test dies on a lossy re-encode, while the ones that survive transport are coarser.
| Method | What it reads | Reported figure | Survives compression |
|---|---|---|---|
| Sample-domain | Waveform discontinuity | 0.965 true-accept at zero false accepts (Korycki, 2013) | No, uncompressed only |
| Environmental signature | Background bed across the seam | Localizes after MP3 at 32 kilobits per second and above (Zhao et al., 2017) | Yes |
| ENF phase | Break in the mains hum | About 6 percent equal-error on a single edit (Nicolalde-Rodriguez et al., 2010) | Partly |
Reach and precision pull against each other, so no one method covers every file. Machine-learning detectors sit outside this table on purpose: they score well inside their training distribution and markedly worse on unfamiliar datasets, which is why cross-condition validation matters more than a single headline accuracy.
Why copy-move has a sharper test
Copy-move is the one edit that comes close to a demonstrable finding, because a segment copied from the same file is provable by signal subtraction: if two intervals cancel to near-silence when one is subtracted from the other, one was copied from the other. The ENFSI manual prescribes exactly this, naming signal subtraction as the verification step, with confirmation by ear made mandatory (ENFSI, 2022). Automated detectors find the candidate pairs first, from 96.59 percent (Imran and colleagues, 2017), through precision and recall of 0.978 and 0.981 (Maksimovic, Aichroth and Cuccovillo, 2021), to 99.70 percent (Yang, Liu and Cao, 2023), and the robust pitch-and-formant method of Yan, Yang and Huang (2019) holds accuracy under added noise and MP3. The full method is in copy-move detection in audio.
Where do false positives come from?
The signals that flag an edit also fire on innocent audio. An impulsive noise, a door slam or a dropped microphone can look like a splice discontinuity, and a repeated word, a looped sample or a stretch of applause can look like a copy-move. This is why auditory confirmation is mandatory rather than optional, and why interpretation depends on content: a repeated chorus in a song is normal, a repeated word in a witness statement is not. The finding is a lead to be checked, never a conclusion on its own.
What a re-encode does to the evidence
Compression is the great eraser. A lossy re-encode after the edit destroys the sample-domain discontinuity that the sharpest splice test depends on, so a heavily re-encoded clip pulled off a messaging app can return a clean result simply because too little signal survived to test. Absence of a detected edit is not evidence that no edit occurred. The traces chosen for their reach, ENF phase and environmental continuity, are selected precisely because they survive further than the raw waveform does.
How an examiner reports it
No audio-forensic method outputs the word authentic. Each edit trace is weighed as a strength of support for one proposition over another and reported that way, never as a verdict (ENFSI, 2015), and a finding earns weight only when independent traces agree. Dating the clip by its mains hum and reading its seams on a spectrogram are the companion reads, and the whole picture comes together in what forensics can learn from a recording.
Sources
- Korycki, R. (2013). Time and Spectral Analysis Methods with Machine Learning for the Authentication of Digital Audio Recordings. Forensic Science International 230(1):117-126. DOI: 10.1016/j.forsciint.2013.02.020
- Zhao, Chen, Wang, Malik (2017). Audio Splicing Detection and Localization Using Environmental Signature. Multimedia Tools and Applications 76(12):13897-13927.
- Nicolalde-Rodriguez, Apolinario, Biscainho (2010). Audio Authenticity: Detecting ENF Discontinuity with High Precision Phase Analysis. IEEE Transactions on Information Forensics and Security 5(3):534-543. DOI: 10.1109/TIFS.2010.2051270
- 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
- Maksimovic, 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).