Your voice, kept

What it is

Photographs survive. Voices usually do not. Most families have hundreds of pictures of a grandmother and not one minute of her talking, and the absence is sharper than people expect.

A voice clone is a synthetic model of a person's voice, built from recordings of that person speaking, which can then produce speech in the same timbre and cadence. It is a file, produced by software, from audio. It is not a person, it is not a presence, and it is not a continuation of anybody. It is a way of hearing how someone said things after they have stopped saying them.

MemoriesBox builds one only from recordings the person made themselves, knowingly, while alive and taking part.

How it works

1. The person records. Ordinary sessions of the guided interview, on a phone or a computer, in a normal room. No studio, no special microphone.
2. The audio is kept as audio. Not transcribed and discarded. The recording is the artefact; the transcript is a convenience.
3. The person presses "Create my digital self" in Settings → Voice. Nothing is built until they do. About 30 minutes of recorded speech makes a good model and is the target the page shows; the hard minimum the vendor accepts is far lower, and a clone from a few minutes sounds like a clone from a few minutes.
4. The model is built and held by ElevenLabs, in the United States, using its instant-voice-cloning method: one model per account, from that account's recordings, replaced if it is rebuilt, deleted when the person retires it or deletes the account. It is the one thing in the archive stored outside the EU, and it is said so on /trust/where-your-data-lives. The vendor works under terms that exclude our recordings from training its general models. It is not licensed, sold, or used to train anything of ours either.
5. Later, the one model speaks in all five site languages. A single multilingual voice model reads the digital self's replies, and reads the interview questions aloud; there is no separate per-language model. Every one of those conversations says on screen what it is.

The thirty-minute threshold

Thirty minutes is a small amount of talking. It is two cups of tea. It is roughly the length of one television episode, and it is less than most people spend on the phone about a parcel.

The practical consequence is that the voice is secured early. By the time a family is worrying about whether there will be enough material, there usually already is. The rest of the recording is for the story, not for the voice.

This matters most in the situation nobody wants to plan for. 7.4 million Americans aged 65 and over are living with Alzheimer's in 2026, projected to reach about 13.8 million by 2060 (Alzheimer's Association, 2026 Facts and Figures). After a diagnosis, the first half hour is the part to do this week. What to record, in what order, is set out on /guides/before-dementia.

What it is built from, and what it is never built from

The right-hand column is not hypothetical. Seance AI builds its simulations from the surviving person's description of the person who died, and states on its own site that it is for entertainment purposes. In Arizona in May 2025, an AI video of Chris Pelkey, a murder victim, delivered his own victim-impact statement in court. Those are reconstructions of somebody who cannot be asked whether they agree.

What it will not do

This is the most important section on the page, and it is short on purpose.

• It will not invent opinions the person never expressed. The archive answers from what was actually recorded. If a family asks what he would have thought about something that happened after his death, the honest answer is that he never said, and that is the answer they get. A system that guesses is writing fan fiction about somebody's father.
• It will not build a voice from photographs, messages, or someone else's description. If there are no recordings the person made themselves, there is nothing for us to work from, and we will not make something up to fill the gap. See /trust/consent.
• It requires the person's own recorded, consenting participation. Not a family member's permission on their behalf. Not a checkbox in the Terms agreed by whoever holds the credit card. The person, recorded, taking part, while alive.
• It will not hide what it is. Every conversation carries a line on screen saying it is a digital imitation built only from the person's own recordings, in all five languages, which EU AI Act Article 50 has required since 2 August 2026. Generated audio is stored with metadata marking it as AI-generated, which is the Act's machine-readable marking. The plain-language version is on /trust/eu-ai-act-article-50.
• It will not be sold, licensed, or used to train a general model. The recordings exist to build one family's archive. There is no second use. It is written into the Terms, and the vendor is bound by contract to the same.
• It will not sound perfect. A clone built from thirty minutes of a person with a cold, in a kitchen with a radio on, sounds like that. More recording makes it better. Nothing makes it flawless.
• It will not be handed over as a keepsake file. A model is not a document, and it may sit with a third-party vendor under terms that do not allow transfer. The durable inheritance is the audio and the transcripts, which is why /features/export matters more than this page does.
• It will not chase anyone. No streaks, no push notifications. One emailed question a week at most, to someone who stopped recording, with a one-click opt-out. See /features/companion.
• It will not outlive your decision. Settings → Voice → "Retire my digital self" deletes the model at the vendor and the character brief with it. The recordings stay.

What it's for

The reason people give is almost never the one they start with. They say they want the stories. What they describe afterwards is a voicemail they cannot delete, or a video where somebody laughs off-camera for two seconds.

47% of Americans regret not recording or documenting a conversation with someone who has since died, and 59% have never done it at all (YouGov, June 2022, n=6,090). That survey covers recording or documenting, so it is not a measurement of voice specifically. It is still the clearest statement of the problem available: the intention is nearly universal and the follow-through is not.

Of the eleven life-story services we compared in August 2026, none clones the storyteller's real voice. Several keep audio, which is a good thing and worth having on its own. Two AI-legacy products build clones: Avataari, a Melbourne company that does not disclose its hosting region, and Afterlife AI, operated from New South Wales. MemoriesBox is the only one of them hosted in the European Union. The comparison that proves it, with prices verified on 17 August 2026, is on /compare.

Questions people ask

How much recording do you need to clone a voice?

About 30 minutes for a good one, which is usually two or three ordinary sessions and is reached long before the life story itself is finished. The vendor will build a rough clone from far less; the settings page shows the 30-minute target because that is the point where it starts to sound like the person rather than like a clone.

Can you clone the voice of someone who has already died?

Not unless they recorded with MemoriesBox while they were alive. We do not build a voice from photographs, messages, video found online, or a relative's description. If there are no recordings the person made themselves, the answer is no, and it stays no.

Is voice cloning legal in Europe?

Yes, with conditions, and the conditions are the interesting part. A voice used to identify a person can fall under GDPR Article 9, which requires explicit consent, and EU AI Act Article 50 has required disclosure and machine-readable marking since 2 August 2026. The detail is on /trust/is-voice-cloning-legal-in-europe.

Who owns the voice clone?

Nobody owns a voice. You hold the recordings and can export them at any time, you decide who gets access, and the model built from them is held for your account and licensed to no one. The four things people mean by ownership are separated out on /trust/who-owns-a-voice-clone.

---

Related: What MemoriesBox does · What the avatar does with the voice · Consent, and who agreed to be recorded · Where your recordings live · Is voice cloning legal in Europe · The full comparison table

Thirty minutes is the whole threshold. Start here. $139 a year, first 14 days free, cancel anytime.