VoiceBlur protection levels

Last updated: 9 November 2026

Every voice you try in VoiceBlur gets a protection level. A level says who the recording holds up against, from a casual listener to anyone at all. It counts only if every level below it holds too.

Level 1 · Masked

Holds up against: People who know the speaker, listening casually.

What it means
The voice itself, its timbre, is replaced. A listener no longer hears the speaker. Most voice-changer apps stop here: in our lab a strong pitch and formant shift still left speaker-recognition software at 9 % error, so it found the speaker almost every time.
Does not cover
Speaker-recognition software. Words, accent and manner of speaking.
Standards
Perceptual masking. Under the GDPR the recording is still personal data.
VoiceBlur
Every voice reaches this.

Level 2 · De-identified

Holds up against: Speaker-recognition software, given a recording of the speaker's real voice.

What it means
A speaker-recognition model compares the blurred voice with the original and does not call them the same person. VoiceBlur runs this check on your phone for every voice sample, on your own recording.
Does not cover
An attacker with VoiceBlur and recordings of a few suspects. Words, accent, background sounds.
Standards
VoicePrivacy Challenge “ignorant” attacker (original voice, stock recognizer). ISO/IEC 20889 de-identification. Under the GDPR still personal data, at most pseudonymised (Art. 4(5)): re-identification is possible with additional information.
VoiceBlur
Checked for each voice sample on your phone. Most voices pass with a clear margin; if one does not, the app says so and you try another.

Level 3 · Unlinkable

Holds up against: Someone who has VoiceBlur too and blurs recordings of suspects the same way.

What it means
Even an attacker who knows exactly how the recording was made cannot link it to the speaker better than chance. That needs the speaker's timing and intonation regenerated too.
Does not cover
What is said, and who could know it.
Standards
VoicePrivacy Challenge “lazy-informed” and “semi-informed” attackers, error near the 50 % of chance. Unlinkability as in ISO/IEC 24745 and ISO/IEC 30136. The linkability criterion of the Article 29 Working Party (Opinion 05/2014).
VoiceBlur
Not reached. In the lab, an attacker with VoiceBlur picked the right speaker out of 10 suspects in 57–74 % of cases (chance is 10 %). VoiceBlur keeps the speaker's timing so video stays lip-synced; that timing is what gives it away.

Level 4 · Anonymous

Holds up against: Anyone, by any means reasonably likely to be used, including what is said.

What it means
Nothing in the recording, not the voice, the words, the accent, the background or the context, lets anyone single the speaker out, link them or infer who they are.
Does not cover
Nothing, by definition. Which is why no tool can promise it.
Standards
Anonymous information under GDPR recital 26. All three criteria of Opinion 05/2014: no singling out, no linkability, no inference. The UK ICO’s “motivated intruder” test.
VoiceBlur
Beyond any voice tool. It depends on what is said and who could know it. See the checklist below.

How the check works

For each voice sample, a speaker-recognition model (ECAPA-TDNN) on your phone compares the first 10 seconds of your recording with the same 10 seconds in the new voice. It is the same comparison an attacker with a recording of you would make, and nothing leaves the phone.

The match line is set where only 1 in 100 pairs of different people would be mistaken for the same person. It was calibrated on 40 LibriSpeech speakers × 16 voices. In that test the phone's score predicted what an attacker with a different recording would see (correlation 0.89): under the app's clear-margin line, that attacker linked none of 545 samples; just under the match line, about 1 in 11; over it, more than half. Each voice sample is shown in one of those three zones.

Under the VoicePrivacy Challenge protocol, where every file gets a random voice, the lab measured 45 % error, close to the 50 % of chance. VoiceBlur has 16 voices, so a determined attacker can try all of them. That is the harder test the app grades against.

Which level do I need?

Sharing a story, a podcast or a clip where nobody is working hard to find out who you are: Level 2 is what you want. Pick a voice with a clear margin.

When a specific organisation could record the people it suspects and run them through VoiceBlur: Level 2 is not enough on its own. Keep the group of people who could have said it large, take identifying details out, and use a channel built for confidential reporting. In the EU, the Whistleblower Directive (2019/1937) entitles you to one.

Before you share

A new voice does not hide what is said. Listen once and check:

Sources