Voice cloning explained simply: how AI copies voices, what is allowed and how to protect yourself

What a voice clone is and how it is made

In voice cloning, an AI analyses recordings of a person and captures what makes their voice distinctive, such as pitch, speaking pace, intonation, timbre and accent. It stores these features as a kind of numerical profile. It can then speak any typed text as if the person were speaking. The provider ElevenLabs describes six steps: gathering recordings, cleaning them, extracting features, training the model, generating new speech and deploying the clone.

Germany's Federal Office for Information Security (BSI) distinguishes two methods. With text-to-speech, the AI reads out a given text in the imitated voice. With voice conversion, someone speaks themselves and the AI turns their voice into that of the target person.

How much material is needed depends on the quality wanted. ElevenLabs says a few minutes of audio for a quick clone and two to three hours for a professional clone. The BSI describes research approaches that manage with just a few seconds, though at the expense of quality.

Legitimate uses of voice clones

Voice clones are not just a tool for fraudsters. Apple offers a feature called Personal Voice on iPhone, iPad and Mac: you record sentences and can then have typed text spoken in a synthetic voice that sounds like yours, for example via Live Speech or augmentative communication apps. According to Apple, the recordings are processed securely on the device. You need at least an iPhone 12 with iOS 17. An important limitation: according to Apple, German is not supported; Personal Voice is only available in English (United States), Mandarin Chinese (China mainland) and Spanish (Mexico).

Commercial providers such as ElevenLabs offer voice clones for audiobooks, videos and voice assistants. ElevenLabs describes its own safeguards: anyone cloning a voice must confirm that it is their own or that they have explicit rights to it. According to ElevenLabs, you may only create a professional clone of your own voice, and even the other person's consent is not enough; a verification step checks that the voice really belongs to you. If someone wants to share their voice with you, they create and verify the clone in their own account and then share it with you. According to the provider, voices of celebrities and public figures are blocked. These are the provider's own statements; whether they work without gaps cannot be verified from outside.

How AI voices made their way from the lab into hotlines and audiobooks is told in our article on voice AI.

What applies in Germany and the EU

Unlike the right to one's own image, the right to one's own voice is not set out in a specific law in Germany. However, on 20 August 2025 the Berlin Regional Court II ruled that the general right of personality also protects a person's voice (case no. 2 O 202/24). A YouTuber had used, without consent, an AI voice resembling the well-known dubbing voice of a voice actor. What mattered to the court was that the audience recognised the voice and could assume the actor had agreed. The YouTuber had already given an out-of-court undertaking to stop in 2023. The court awarded the actor 4,000 euros as a notional licence fee and 1,155.80 euros in legal costs, each plus interest.

On 7 October 2026, the German federal cabinet also adopted a draft law against digital violence. Among other things, it provides for new criminal offences for the unauthorised distribution of reputation-damaging deepfakes and for identity misuse. The Bundestag and Bundesrat still have to approve it; the Federal Ministry of Justice expects the rules could apply from mid-2027. According to LTO, anyone who obtains money using someone else's identity can already be prosecuted today, for example for fraud.

Across the EU, Article 50 of the AI Act has applied since 2 August 2026: AI providers must mark generated audio in a machine-readable way. For systems already on the market before 2 August 2026, the European Commission says this marking obligation only applies from 2 December 2026. Anyone publishing a deceptively realistic voice fake in a professional or business capacity must disclose it as AI-generated. Purely personal use, for example sharing a deepfake on social media as a private individual, falls outside the AI Act according to the Commission; other rules such as personality rights still apply. Our assessment: this does little against scam calls, because criminals do not follow labelling rules.

How to protect yourself and your family

In January 2026, the consumer advice centre in Brandenburg warned that fraudsters take short voice snippets from social media, use them to create a relative's voice and fake an emergency on the phone. Such calls are hard to recognise as fakes; unnatural conversation flow or gaps can be clues. The BSI also lists a metallic sound, mispronunciations, monotonous intonation and delays before answers. You should not rely on such signs, though.

The German police advise the following for shock calls:

Hang up immediately. That is not rude; it gives you time to breathe.

Call the person themselves, or someone you trust, on a number you know.

Agree on a code word with your family and friends so you can recognise each other on the phone.

Never give details of your finances and never hand over money or valuables to strangers.

If in doubt, call 110, the German police emergency number, and always go to the police after a fraud.

The consumer advice centre also recommends asking questions only the real person can answer and, after a suspicious call, noting the date, time, number and course of the conversation. As a precaution, the police advise shortening or deleting your entry in the phone book. How to spot AI fakes in images, videos and calls in general is explained in a separate article.