Contents
An audio recording carries two forensic traces beyond the words, the electric network frequency, a mains-hum stamp of time and power grid, and the microphone fingerprint, a device signature written into the frequency response, and because both live in the signal rather than the metadata, a tag strip does nothing to them and removing either one cleanly is hard. This article explains what they are, what they reveal and why the verdict is that you can degrade them at a cost but not delete them without leaving a mark.
Two fingerprints live in the sound, not the metadata
When you clear the tags on an audio file you remove the filename, the timestamps and any embedded device fields. You do not touch the waveform. Two identifying traces live inside that waveform, and neither is metadata. The first is a record of the mains power around the recorder. The second is the acoustic signature of the microphone and device that captured it. Both are the audio equivalent of a camera’s sensor fingerprint: content, not container, and therefore untouched by a metadata strip. The image-side version of this problem is in can a photo be traced without metadata.
ENF: a mains-hum stamp of when and which grid
Mains electricity hums at a nominal 50 or 60 hertz, and that frequency drifts slightly, moment to moment, in a way shared across an entire power grid. Any recorder plugged into the mains, or simply near it, picks up that drifting hum. Grigoras (2005) set out the founding electric network frequency criterion at the IAFPA meeting: because the whole interconnected grid carries the same fluctuating frequency, the captured hum is a shared timestamp that can be checked against a reference log of the grid.
What it reveals is twofold. Cooper (2008), working in the Metropolitan Police forensic audio context, showed that ENF lets an examiner ascertain the date and time of a recording and detect edits, and noted that even battery-powered devices pick the signal up through the electromagnetic field near the mains. Hajj-Ahmad, Garg and Wu (IEEE TIFS 2015) add place: ENF “is a signature of power distribution networks that can be captured by multimedia signals recorded near electrical activities,” and its statistics can infer which grid the recording was made on. In practice that names the synchronous grid, continental Europe, the United Kingdom, North America, rather than the street.
The microphone fingerprint: a device signature in the frequency response
Every microphone and its surrounding electronics colour the sound in a slightly individual way, imposing a frequency response that is consistent enough to identify the device. Buchholz, Kraetzer and Dittmann (IH 2009) classified microphones from their recordings using Fourier coefficients and reported up to 93.5 percent correct classification across seven microphones. Hanilci, Ertas, Ertas and Eskidere (IEEE TIFS 2012) took it to consumer devices, reaching closed-set identification rates of 92.56 and 96.42 percent on a set of 14 different cell phones, because each device, in their words, “introduce[s] a convolutional distortion on the input speech, hence leaving its own tell-tale footprints.” The signature rides in the speech itself, not in any separate channel.
Why stripping either is hard
This is where the real limits sit, and why a clean strip is difficult.
ENF removal usually means notching out the 50 or 60 hertz line. The trouble is that the nominal frequency is stable, 60 hertz in the Americas and 50 hertz in most other regions, varying only between 59.9 and 60.1 hertz, so a notch has to be narrow, and narrow notching leaves higher harmonics of the hum behind. Worse, it replaces the natural mains band with an unnaturally clean gap, which is itself a flag. Chuang, Garg and Wu (IEEE TIFS 2013) studied precisely this, describing operations that “remove and alter the ENF signal while trying to preserve the host signal, and devises detection methods targeting these operations.” In other words, the research on removing ENF comes packaged with research on detecting the removal. It is also the one trace here with a court-facing best-practice standard behind it.
The microphone fingerprint is harder still, for a different reason. It is not noise you can denoise away, it is colour woven through the speech spectrum. Cuccovillo, Giganti and Bestagini (ACM ICMR 2022) show that denoising a clip does not erase the device signature, it sharpens it, reporting “an average accuracy increase of about 25%” for microphone classification on denoised audio. So the obvious cleanup step makes the fingerprint easier to read, not harder. Genuinely suppressing it means whitening the spectral envelope, which audibly degrades the voice.
Where the ceiling sits
The realistic position mirrors the image sensor fingerprint exactly. Both ENF and the microphone signature can be degraded, by narrow notching and filtering for the first, by spectral whitening for the second, and both degradations cost audio quality. Neither can be deleted cleanly, and the attempt tends to be detectable: an over-clean mains band and the detection methods built alongside the removal research for ENF, and the fact that ordinary denoising strengthens rather than removes the microphone trace. Treat these as forensic realities to understand, not switches to flip. If hiding your speaker identity is the actual goal, that is a different and better-defined task, covered in voice anonymizer tools that work.
Sources
- Grigoras (2005). Digital Audio Recording Analysis: The Electric Network Frequency Criterion. IAFPA.
- Cooper (2008). The Electric Network Frequency (ENF) as an Aid to Authenticating Forensic Digital Audio Recordings. AES 33rd International Conference.
- Hajj-Ahmad, Garg, Wu (2015). ENF-Based Region-of-Recording Identification for Media Signals. IEEE TIFS.
- Chuang, Garg, Wu (2013). Anti-Forensics and Countermeasures of Electrical Network Frequency Analysis. IEEE TIFS.
- Buchholz, Kraetzer, Dittmann (2009). Microphone Classification Using Fourier Coefficients. IH 2009.
- Hanilci, Ertas, Ertas, Eskidere (2012). Recognition of Brand and Models of Cell-Phones From Recorded Speech Signals. IEEE TIFS.
- Cuccovillo, Giganti, Bestagini (2022). Spectral Denoising for Microphone Classification. ACM ICMR 2022.