Contents
Stripping the metadata from a voice recording clears the same container tags as any audio file, the filename, the timestamps and the device fields, but it leaves the speaker’s voiceprint and the microphone signature intact in the audio, so the recording can still be tied to a speaker and to a device after the tags are gone.
The tag strip is the same as any audio file
A voice recording is an audio file, so its metadata lives in the same place and comes off the same way. An MP3 keeps its fields in an ID3 tag, an M4A in MP4 atoms, a WAV in RIFF INFO or Broadcast WAV chunks, and a FLAC or Opus file in a comment header. Those fields hold the filename, the capture timestamp, and sometimes a device string or an account name, and some voice-memo and call-recording apps add a custom note title or an owner field on top. ExifTool clears the whole block in one directive, ffmpeg drops it on a stream copy, and mutagen edits the frames directly. It is worth doing, it takes seconds, and it is worth confirming the entire block is gone rather than just the obvious title.
One caveat carries over from any audio file: most platforms strip the container metadata from the public copy on upload, which is easy to mistake for full protection. The platform may have logged the original fields before stripping them, and either way the audio it keeps is unchanged. Deleting the tags rewrites the container, not the content, and for a voice recording the content is where the identity is, the same lesson remove metadata from an audio or MP3 file draws for audio generally.
What survives is the voice and the device
Here is the catch. Two identifiers stay in a voice recording after every tag is gone, and neither is something a tag strip can reach.
The first is the voiceprint. Speaker-recognition systems do not read the filename, they read the voice, building a numerical model of how a particular person sounds and matching it against known samples. Research treats a voiceprint as biometric, an identifier of the person rather than of the file, which is why it is unaffected by anything you do to the metadata. Wipe every tag and re-encode the file, and the same voice is still there to be matched. The label changes; the speaker does not.
The second is the microphone fingerprint. Every device colours the sound in an individual way, a frequency response consistent enough to name the specific microphone. Buchholz, Kraetzer and Dittmann (IH 2009) classified microphones from recordings using Fourier coefficients at up to 93.5 percent accuracy across seven microphones, and Hanilci, Ertas, Ertas and Eskidere (IEEE TIFS 2012) reached 92.56 and 96.42 percent identification on 14 cell phones, because each device “introduce[s] a convolutional distortion on the input speech, hence leaving its own tell-tale footprints.” More recent transformer-based methods from Qamhan, Alotaibi and Selouani (Applied Sciences 2023) report identification from 97.6 to 99.98 percent inter-model accuracy. Cleaning the clip up does not help: Cuccovillo, Giganti and Bestagini (ACM ICMR 2022) found that denoising yields “an average accuracy increase of about 25%” for microphone classification, so the tidy-up step sharpens the signature rather than hiding it. The full account is in strip ENF and microphone fingerprint from audio.
If hiding the speaker is the real goal
Removing metadata from a voice recording is a narrow win. It is worth doing as hygiene, and it clears the fields that name the file and the account, but it is not anonymisation of the speaker, because the voiceprint and the device signature ride in the sound the file contains. The tag is the label; the voice is the evidence.
If what you actually want is to keep the recording from being matched to you as a speaker, that is a content-level task and a different one. Whether it can be done at all, and how well, is covered in does voice anonymization actually work, and the practical tool options are in voice anonymizer tools that work.
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.
- Qamhan, Alotaibi, Selouani (2023). Source Microphone Identification Using Swin Transformer. Applied Sciences.
- Cuccovillo, Giganti, Bestagini (2022). Spectral Denoising for Microphone Classification. ACM ICMR 2022.