Voice biometrics uses the characteristics of a person's voice to confirm or establish who they are. Under GDPR, voice data processed for the purpose of uniquely identifying someone is biometric data, a special category under Article 9, which in commercial settings usually means explicit consent is required. Transcribing speech is a different activity: it processes what was said, not who said it, and does not by itself make audio biometric data.
Key takeaways
- Speech recognition and voice recognition are different tasks with different legal treatment. The purpose of processing decides whether voice data is biometric.
- Voice authentication in the EU generally rests on explicit consent under Article 9(2)(a), with a real alternative for people who decline.
- Verification (is this the claimed person?) and identification (who is this?) carry different risks, and the AI Act treats them differently.
- Voice cloning has weakened voice-only authentication. Spoofing detection and a second factor are now standard advice.
- Voiceprints must be deleted when consent is withdrawn, and a DPIA is usually required before deployment.
How voice biometrics works#
A speaker recognition system maps a stretch of speech to a fixed-length vector, a speaker embedding, that captures vocal characteristics while ignoring the words spoken. Earlier systems used i-vectors; modern ones use neural embeddings such as x-vectors and their successors. Two recordings of the same person produce embeddings that are close together; recordings of different people produce embeddings that are further apart.
Deployment has two phases. At enrolment, the system records one or more samples and stores the resulting embedding as the person's voiceprint. At verification, a new sample is embedded and compared with the stored voiceprint, and the similarity score is tested against a threshold.
The threshold sets the balance between two errors. A false accept lets an impostor in; a false reject turns away the genuine user. The equal error rate, where the two are equal, is the usual single-number summary in research, but deployments choose an operating point based on which error costs more.
Verification, identification and diarization#
| Task | Question | Links voice to a named person? |
|---|---|---|
| Speech recognition | What was said? | No |
| Diarization | Which segments share a speaker? | Not by itself; labels are anonymous |
| Verification (1:1) | Is this the person they claim to be? | Yes, against one enrolled voiceprint |
| Identification (1:N) | Who, among enrolled people, is this? | Yes, by searching many voiceprints |
Diarization uses speaker embeddings too, but it labels speakers as "A" and "B" within one recording. Whether it amounts to biometric processing depends on what happens next: embeddings discarded after the job are very different from embeddings stored and matched across calls. See diarization vs identification vs verification.
GDPR: when voice is biometric data#
Article 4(14) GDPR defines biometric data as personal data resulting from specific technical processing relating to the physical, physiological or behavioural characteristics of a person, which allow or confirm their unique identification. Article 9(1) prohibits processing biometric data for the purpose of uniquely identifying a natural person, unless one of the exceptions in Article 9(2) applies.
The purpose test matters. A call recording is personal data, but it is not special category biometric data merely because a voice could in principle be identified from it. It becomes biometric data in the Article 9 sense when it is processed by technical means to identify or authenticate the speaker. Recital 51 makes the same point for photographs. Transcription, which extracts words, does not meet that test; building and matching voiceprints does. The broader status of recordings is covered in GDPR and voice data.
Consent#
For voice authentication offered by a bank, telecom operator or public service, the realistic legal basis is usually explicit consent under Article 9(2)(a). Consent must be freely given, specific, informed and unambiguous (Article 4(11)), must be as easy to withdraw as to give (Article 7(3)), and for special category data must be explicit. In practice that means:
- a clear, separate request before enrolment, not a clause in general terms;
- an alternative way to authenticate, such as security questions or an app, so that refusing is a real choice;
- a simple way to withdraw, after which the voiceprint is deleted;
- care in employment settings, where the imbalance of power between employer and employee makes freely given consent hard to establish (Recital 43; EDPB Guidelines 05/2020).
Article 9(4) allows member states to add further conditions for biometric data, so national law needs checking as well. Large-scale processing of special category data is a listed case for a data protection impact assessment under Article 35(3)(b).
The EU AI Act#
Regulation (EU) 2024/1689 lists remote biometric identification systems as high-risk in Annex III, but excludes biometric verification whose sole purpose is to confirm that a person is who they claim to be. A caller authenticating against their own voiceprint is verification; searching a voice against a database is identification. The Act also prohibits biometric categorisation that infers sensitive attributes such as political opinions or sexual orientation (Article 5(1)(g)), and restricts emotion recognition from voice. See the EU AI Act and speech recognition.
Spoofing and voice cloning#
A voice is not a secret. It can be recorded and replayed, and modern text-to-speech and voice conversion can imitate a specific speaker from a short sample. Attacks fall into three groups:
- Replay: playing a recording of the genuine speaker.
- Synthesis: generating the target voice from text with a cloned model.
- Conversion: transforming an impostor's live speech to sound like the target.
The ASVspoof challenge series, running since 2015, provides shared datasets and metrics for spoofing countermeasures, and ISO/IEC 30107 defines the framework for presentation attack detection in biometrics. Countermeasures help, but detection trained on known synthesis methods can fail on new ones.
Outside the EU#
Other jurisdictions regulate voiceprints specifically. The best-known example is the Illinois Biometric Information Privacy Act (2008), which lists voiceprints as biometric identifiers, requires written informed consent before collection and a published retention schedule, and gives individuals a private right of action. Organisations operating in both markets often design to the stricter of the two regimes.
How Reson8 approaches this#
Reson8's products are speech-to-text services: real-time streaming, batch transcription and the Turns endpoint for voice agents. They extract what was said. Audio is not retained and customer data is not used for model training, so recordings sent for transcription do not accumulate into a store from which voiceprints could later be built.
Common questions#
Is every call recording biometric data?
No. Under GDPR, voice data becomes special category biometric data when processed to uniquely identify a person. A recording kept for quality purposes is personal data but not, on that basis alone, biometric data.
Can I use legitimate interest instead of consent for voice authentication?
Legitimate interest is not an Article 9 exception. You need one of the Article 9(2) conditions, and for commercial voice authentication that is usually explicit consent, subject to national law.
How long can a voiceprint be kept?
Only as long as necessary for the purpose, and it must be deleted when the person withdraws consent or closes their account.
Is voice authentication still safe with voice cloning?
On its own, less than it was. Used with spoofing detection and a second factor it remains useful, particularly for reducing friction in low-risk interactions.
This article is general information, not legal advice.
References#
- Regulation (EU) 2016/679 (General Data Protection Regulation), Articles 4, 7, 9 and 35. EUR-Lex.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5 and Annex III. EUR-Lex.
- European Data Protection Board. Guidelines 05/2020 on consent under Regulation 2016/679. EDPB.
- Snyder, D., Garcia-Romero, D., Sell, G., Povey, D., & Khudanpur, S. (2018). X-vectors: Robust DNN embeddings for speaker recognition. ICASSP.
- Dehak, N., Kenny, P., Dehak, R., Dumouchel, P., & Ouellet, P. (2011). Front-end factor analysis for speaker verification. IEEE Transactions on Audio, Speech, and Language Processing.
- ASVspoof: automatic speaker verification spoofing and countermeasures challenge. asvspoof.org.
- ISO/IEC 30107-1:2016. Information technology: Biometric presentation attack detection: Framework.
- Illinois Biometric Information Privacy Act, 740 ILCS 14 (2008).