Contents
Removing the metadata from an audio file is the easy and worthwhile half of the job: the ID3 tags on an MP3, or the atom fields inside an MP4 or M4A, strip out in seconds with a standard tool, which deletes the filename, timestamps and device fields that are the fastest way to place a recording. What a tag strip does not touch are the two identifying traces that live in the sound itself, the microphone fingerprint and the mains-hum ENF stamp, so clearing the tags is hygiene, not anonymity.
What a tag strip actually removes
Every audio file ships with a block of structured text alongside the samples. On an MP3 that block is the ID3 tag, in one of its versions, ID3v1, ID3v2.3 or ID3v2.4. On an MP4 or M4A it is a set of atoms such as ©nam, ©ART and ©day held inside the udta and ilst structure. A WAV carries RIFF INFO chunks, or a bext chunk if it is a Broadcast WAV. A FLAC or Opus file keeps a Vorbis-style comment header. All of these are container fields. They hold the title, artist, album, year, encoder and sometimes an owner name or a device string, and they are fully removable.
Clearing them takes one step with a standard public tool. ExifTool wipes the whole tag block with a single all-tags directive. ffmpeg drops the container metadata on a stream copy. The Python library mutagen edits or deletes the frames directly. A plain container re-mux or a re-encode also discards most of the tag block as a side effect. None of this is exotic, and it is worth doing before you share anything.
The catch is the same one that applies to a photo’s EXIF: the tag block is separate from the audio. Deleting it rewrites the container, not the content. Everything encoded in the samples is left exactly as it was.
What survives lives in the signal
Here is the catch. Research on audio forensics shows that the traces that actually identify a recording are not in the tag block at all, they are in the waveform. The tag is the label; the signal is the evidence. A metadata strip removes the label and leaves the evidence untouched.
The first surviving trace is the microphone fingerprint. Every microphone and its surrounding electronics colour the sound in a slightly individual way, a frequency-response signature consistent enough to identify the specific 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 the same idea to consumer handsets, reaching 92.56 and 96.42 percent closed-set identification on 14 cell phones, because each device, in their words, “introduce[s] a convolutional distortion on the input speech, hence leaving its own tell-tale footprints.” Newer work pushes the ceiling higher: Qamhan, Alotaibi and Selouani (Applied Sciences 2023) report source-microphone identification from 97.6 to 99.98 percent inter-model accuracy. None of that reads a tag. It reads the sound.
The obvious cleanup step makes this worse, not better. 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. Running a recording through noise reduction to tidy it up hands the identifier a cleaner read.
The second surviving trace is the electric network frequency, the faint mains hum at a nominal 50 or 60 hertz that a recorder picks up on or near mains power. It drifts moment to moment in a way shared across a whole grid, so it acts as a time-and-place stamp, and it too rides in the waveform where a tag strip cannot reach it. The full account of both signals, and how far each can be degraded, is in strip ENF and microphone fingerprint from audio.
The label is not the evidence
This is the model to carry away. Metadata is a label attached to the file, and you can peel the label off in seconds. The microphone fingerprint and the ENF hum are properties of the sound the file contains, and they stay put when the label is gone. That is why a tag strip is genuinely useful and genuinely limited at the same time: useful because it removes the single quickest way to date and place a recording, limited because it is silent about the identifiers baked into the audio.
One more caveat sits on top. Most platforms strip container metadata from the public copy on upload, which is easy to mistake for full protection. It is not. The platform may already have logged the original fields on its own servers before stripping the copy you can see, and the audio it keeps is unchanged, so the signal-domain traces are still there to read.
Where to go next
Clear the tags first: it is the right first step and it costs nothing. Then be honest about what is left. If your real goal is to stop a recording being tied to you as a speaker, that is a different and better-defined task, covered in voice anonymizer tools that work. If you want to understand the device and grid signatures in depth, see strip ENF and microphone fingerprint from audio. And if the question is whether a file has been edited or tampered with, see audio tampering detection.
Sources
- 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.
- Qamhan, Alotaibi, Selouani (2023). Source Microphone Identification Using Swin Transformer. Applied Sciences.