AI voices in learning: speech synthesis and consent

Last reviewed: October 16, 2026

Short answer

Synthetic voices turn text into natural-sounding speech and make narrated courses quick to produce and easy to change in many languages. Recreating a real person’s voice requires their explicit, documented and revocable consent; a voice is personal data, and AI-generated voices should be recognisable as such to listeners.

Key points

Two kinds: ready-made synthetic voices and recreations of real voices.
A voice is personal data; recreating it requires consent.
Document consent, define the purpose, make revocation take effect immediately.
Label AI-generated voices, especially when they imitate real people.
A figure should never speak with another person’s voice.

How speech synthesis works

Modern speech synthesis turns text into an audio track that is barely distinguishable from a human recording. It takes intonation, stress and pauses into account and speaks many languages. For learning platforms this is a big step forward: courses can be narrated without booking speakers, changes only require regenerating the affected sentence, and further languages are created from the translated text.

Besides ready-made voices, a particular voice can be recreated. From a short voice sample, the system learns how a person sounds and can then speak any text in that voice. A lecturer can thus have their courses narrated in their own voice without recording every sentence themselves.

Uses in learning

The table shows typical uses and what to watch for in each case.

Use	Benefit	What to watch for
Narrating slides	Fast, changeable, multilingual	Pronunciation of technical terms
Avatar as speaker	Visible presenter without filming	Labelling, voice suits the figure
The lecturer’s own voice	Personal, familiar	Consent, revocation
Audio version, podcast	Learning on the move	Natural flow of speech
Read-aloud for accessibility	Access for people with reading difficulties	Understandable pace

A voice recording is personal data under the GDPR. If it is processed to uniquely identify a person, it can fall under the special categories of Article 9 as biometric data. For recreating a voice in learning courses, the person’s explicit consent is in practice the appropriate basis. It must be given freely, on an informed basis and for a specific purpose, and under Article 7 it can be withdrawn at any time.

In employment, voluntariness is a sensitive point because employees may feel under pressure. It should therefore be possible to refuse consent without disadvantage, and there should be an alternative, such as a ready-made voice.

Article 50 of the EU AI Act requires, among other things, that AI-generated or manipulated image, audio and video content that deceptively resembles real people be disclosed. For learning platforms, visible labelling of the AI voice is the safe route.

Consent is only as good as its implementation. It should be in place before any processing, documented and technically effective: without consent no voice is built, and after revocation the voice no longer speaks.

Consent before recording, with name, date, purpose and period of validity
Separate consent for image and voice
Store a signed record
Limit the purpose, for example to the organisation’s own courses
Revocation possible at any time, effective immediately
No use for training third-party models without separate agreement
Offer a ready-made voice as an alternative where there is no consent

Preventing misuse

Recreated voices can be misused, for example to put words in someone’s mouth that they never said. Learning platforms should therefore have clear rules: a voice belongs to exactly one figure, a figure speaks only with its own voice, and there is no automatic substitute voice that steps in when its own is missing. A figure whose voice is unavailable should rather stay silent and show the answer as text.

Equally important is who may create and use voices. Only authorised people should be able to approve figures and voices, and every use should be traceable.

How Universe does it

In Universe, a custom figure is only built once consent has been recorded with name, date, period of validity and a signed record, separately for image and voice. Revocation takes effect immediately: the figure no longer speaks. Every face speaks only with its own voice; there is no substitute voice, and this rule is protected in the code by automated guards. For “My avatar”, about 30 seconds of voice sample is enough; image and voice are processed in the EU and not used to train third-party models. Everything a figure says is labelled as AI-generated.

Terms

Speech synthesis: Generating spoken language from text by software.
Voice clone: A recreation of a specific person’s voice from a voice sample.
Consent: A freely given, informed and unambiguous agreement to a specific processing of data.

Frequently asked questions

Is a voice clone allowed under the GDPR?

With a valid legal basis, in practice usually the person’s explicit consent, and with due regard to purpose limitation, security and the right to withdraw. No voice should be recreated without consent.

Can an employer require an employee to have their voice cloned?

Consent must be voluntary; in employment this needs to be examined particularly carefully. Employees should be able to refuse without disadvantage, and there should be an alternative.

Do AI voices have to be labelled?

The AI Act contains transparency obligations for certain AI-generated content, particularly where it deceptively imitates real people. Labelling is advisable in learning courses in any case.

What happens after consent is withdrawn?

The voice may no longer be used. Good systems implement this technically and immediately, so that the figure no longer speaks with that voice.

How long does a voice sample for a clone need to be?

That depends on the method; some systems manage with short samples of under a minute. Quality matters more than length: a quiet room, a steady distance from the microphone and natural speech without long pauses.

How it works in Universe

Sources