Preservation · Updated 2026-11-16 · 9 min read
Can AI Recreate a Deceased Loved One's Voice? What to Know First
Short answer
AI can imitate a loved one's voice, but not recreate it exactly. Software builds a synthetic voice from clear recordings such as voicemails, voice notes, and videos. It is an imitation made from samples, not the person, and its quality depends on the audio that exists. The originals cannot be replaced, so save those first.
What AI can and cannot recreate from a recording
A synthetic voice, often called a voice clone, is made by software that studies how a person sounded in their recordings, including pitch, accent, pace, and the texture of the voice. It then reads new text aloud in a similar voice. It does not know the person. It reproduces how they sounded, not what they remembered, believed, or would have said.
That is why the careful word for the result is imitation. Whatever a synthetic voice says is text that someone else wrote, and how close it sounds depends heavily on the recordings you start with. The table compares the three things people usually mean when they ask about this.
| Option | What you hear | What it needs | Main limit |
|---|---|---|---|
| Original recordings | Their real voice, exactly as it was captured | Files you already hold, such as voicemails, voice notes, and videos | They cannot say anything new, and they can be lost if one phone or account holds the only copy |
| Synthetic voice | A voice that resembles them, reading text someone writes | Clear recordings of one speaker, and your family's agreement | It is an imitation, only as close as the audio allows, and never the person |
| Written memory | Typed replies shaped by their letters and messages, with no voice | Messages, letters, or a chat export | Software writes the replies from what was saved, so they are never the person's own new words |
Secure the original recordings first
Every option starts with the same material, and it is the one part that cannot be rebuilt. A voicemail deleted by a carrier, or a voice note that disappears with an old phone, cannot be restored, and no software can recover sound that was never kept. If what you want is to hear their real voice again, the original recordings already do that. Save them before you try anything else, and before you close an account or cancel a phone line.
Our voicemail and WhatsApp guides, linked at the end of this page, walk through the export steps. The checklist below applies whichever apps you use.
- Gather every recording in one folder: voicemails, WhatsApp voice notes, voice memos, video calls, home videos, and answering-machine tapes if you have them.
- Copy files instead of moving them. Keep one untouched folder of originals, and do any trimming or uploading from copies.
- Store the originals in two separate places you control, such as a computer drive and an external drive or a relative's backup.
- Do not let a phone carrier, a chat app, or a memorial service be the only place a recording lives.
- Note who is speaking, when it was recorded if you know, and where the file came from. It helps you choose the best clips later.
- Export WhatsApp chats with media attached, and save voicemail files off the phone before the line or the handset changes.
- Tell one other family member where the originals are, so they do not depend on one person's memory or password.
What changes how close a synthetic voice sounds
Researchers score closeness with speaker-similarity tests. RVCBench, a 2026 benchmark of 18 open-source voice-cloning models and 204 speakers, found that background noise in the reference audio hurt many models, and that other people talking over the speaker degraded every model evaluated. The recording you start with largely sets the limit.
The table turns those findings into practical choices. None of it is a promise about any one recording, because every file is different, but it shows where a little care changes the result. Some services help with this step: therefor.us, for example, lists each uploaded file's length and signal level and marks phone-quality recordings.
| Factor | What is known | What to do |
|---|---|---|
| Background noise | In RVCBench, many models lost speaker detail when the reference recording was noisy | Prefer clips recorded in a quiet room over a car, a party, or a street |
| Other voices talking | Every model evaluated lost accuracy as the interfering speech got louder | Choose moments when only one person is speaking |
| Length of the clear clip | Gains were steepest up to roughly 8 to 12 seconds of reference speech, then slowed, though some stronger models kept improving | A few clear seconds can beat a long noisy stretch, and more clean audio can still help |
| Emotion and long passages | RVCBench found open-source models are generally not robust on emotionally expressive prompts or long passages | Expect more variation on emotional or long passages, and treat the result as an approximation |
| Phone-line voicemail | Voicemails are often stored in a narrow, compressed format, so they can sound thinner than voice notes or video | If you hold both, start with the voice note or the video, and still save the voicemail |
| How scores are read | On a widely used speaker-similarity benchmark, two genuine recordings of the same person score only about 0.73 out of 1.0 against each other | Read any score as a guide, and trust your own and your family's listening |
Consent and family: decide together first
No service can confirm what a person who has died would have wanted. Some professional voice-cloning products check that the voice belongs to the person signing up, for example by asking for a live recording or a spoken consent statement. A person who has died cannot give either. Services that do accept uploads from relatives rely on the person uploading to act in good faith.
That makes the decision a family one. Laws about using a person's voice after death differ by country and state, and they mostly concern public or commercial use. This article is general information, not legal advice. For private use, a practical test is whether the people closest to them would be comfortable with it.
A convincing voice can also be used to mislead other people, for example in a message that pretends to come from a relative. Keep synthetic audio private, and if you ever share it, say plainly that it is synthetic.
- Did they ever say, in writing or aloud, what they would want done with their recordings?
- Who are the people closest to them, and have all of them been asked?
- Who may listen, who may add recordings, and who holds the passwords?
- Will anything be shared outside the family, such as on social media or in a video? If so, does everyone agree?
- Would anyone find this upsetting, and how will their wishes be respected?
- What happens to the files if one person changes their mind, or if the service closes?
Privacy and storage: what to ask before you upload
A recording of someone's voice is personal data, and a voice built from it is too. Keep your own copy of everything you upload, so your recordings never depend on one company staying open. Some memorial services have already closed and left families without copies of their own recordings; the article on that is linked at the end of this page.
Work through the four steps below, then use the table to compare any service, including ours, on the same questions. The right-hand column shows what therefor.us states on each point.
- Read the privacy terms, and note anything about sharing, training, or deletion that you cannot find.
- Upload copies, never your only originals.
- Start with one or two clear clips to hear how close the result sounds before adding more.
- Use a strong, unique password, and share access only with people the family has agreed on.
| Question to ask | What therefor.us states |
|---|---|
| Who can hear it, and is it shared? | It is a private memorial space. Nothing is shared, sold, or used to train outside models. |
| Where does the model run? | The persona model is self-hosted, so no outside company holds the family's conversations. |
| Can you hear how close it is before paying? | Every voice gets a match score, measured against what the recordings allow, before anyone pays. |
| Can you keep a copy of the voice? | Voice Preservation includes a downloadable recording of the voice reading text you write. Keep your own originals as well. |
| Does it renew by itself? | No. Plans are prepaid windows, paid in USDT, USDC, or DAI on Ethereum, and nothing renews automatically. |
| Can you remove what it remembers? | Each remembered fact is listed with its source on the person's page and can be removed for good. |
Go at your own pace, and keep the originals either way
Hearing a familiar voice say new words can comfort one person and unsettle another, and the same person may feel both on different days. Neither reaction means you are grieving wrongly. There is no deadline, and nothing here needs deciding today.
A synthetic voice is not a substitute for people who can support you. If grief is making it hard to sleep, eat, or get through the day, a doctor, a counselor, or a bereavement group can help in ways software cannot. If you try a voice and it feels wrong, you can stop, and your original recordings stay exactly as they were.
If you ever feel unsafe, please contact local emergency services or a crisis line right away.
Your options, from keeping recordings to building a voice
You do not have to choose anything now. The options below run from least to most commitment, and the first one is worth doing whatever you decide.
therefor.us shows a built voice as a match score. It measures how close the voice gets compared with the most that person's own recordings allow, so it is not a percentage of likeness to the person. Scores are labeled Strong (80% or more), Good (60 to 79%), Partial (40 to 59%), or Weak (below 40%), and a weak match usually points to short or noisy audio. Prices were checked against the plan list on 2026-11-16, and the pricing page has the current terms.
| Option | What you get | Cost | A good fit when |
|---|---|---|---|
| Keep the recordings only | Originals saved in two places, with no software involved | Free, using your own storage | You are not ready, or you want the audio safe before deciding anything else |
| Written memory (Meet them) | One person, text only, 40 messages, no voice | $0 | You want to see how a written memory feels before hearing a voice |
| Voice Preservation | A voice built from up to 60 minutes of audio for one person, with a downloadable recording of it reading text you write | $39.00 once | You want a recording you keep, without a subscription |
| Remembrance or Family | Voice conversation windows: Remembrance gives one person 60 voice minutes per 30 days, and Family gives five people 300 shared minutes per 30 days | Remembrance $9.99 per 30 days or $89 per year; Family $19.99 per 30 days or $179 per year | You want to hear the voice in conversation and prefer prepaid windows |
Scope note: A synthetic voice is an imitation built from the recordings that exist. It is not the person, it cannot say what they would have said, and no service can restore detail that is missing from the audio. Research figures describe open-source models tested by researchers, not any one family's recordings. This is general information, not legal, medical, or grief-counseling advice.
Common questions
Is there an AI voice for people who have passed away?
Yes, in a limited sense. Some services build a synthetic voice from recordings of someone who has died and read text that a family member writes. It imitates how they sounded and is not the person. Services differ on consent, privacy, and whether you can keep a copy, so ask before uploading anything.
How do you recreate someone's voice who has passed away?
Gather the original recordings and save untouched copies in two places. Choose the clearest clips, with one speaker and little noise, and agree with the people closest to them. Then pick a service whose privacy terms you have read, check how close the result sounds before you pay, and keep the originals regardless.
Can AI recreate a dead person?
No. AI can imitate how someone sounded or wrote, using what they left behind, but it cannot recreate the person, their memories, or what they would have said. Anything it produces is generated from samples and from words that other people wrote.
Can you talk to a dead loved one through AI?
You can exchange messages with software built from their letters, texts, and recordings, but it is software replying, not them. Some people find that comforting and others find it unsettling, and either reaction is normal. On therefor.us, a written memory for one person is free (40 messages, no voice), so you can see how it feels before deciding on anything more.
Can I use AI to make a recording of my loved one who passed?
If you hold clear recordings of their voice, software can read new text in a similar voice and save it as an audio file. Agree on it with close family, label it as synthetic if you ever share it, and keep the originals untouched. Voice Preservation on therefor.us costs $39.00 once and includes a downloadable recording of the voice reading text you write.
Sources
- RVCBench: Benchmarking the Robustness of Voice Cloning Across Modern Audio Generation Models (arXiv:2602.00443)
- Seed-TTS: A Family of High-Quality Versatile Speech Generation Models (arXiv:2406.02430)
- Microsoft Learn: Add voice talent consent for professional voice
- ElevenLabs documentation: Voices
- therefor.us pricing & preservation plans
Start calmly, on your own schedule
Building a written memory for one person is free. When you hold recordings you want preserved, Voice Preservation measures what your clips allow and gives you a file to keep.
Related reading
- Export WhatsApp voice notes before you change phones
- Compare downloadable voice files with subscription memorials
- See the free text tier and Voice Preservation prices
- How to Save Voicemails From Someone You Lost Before Carriers Delete Them
- HereAfter AI Shut Down: How Families Can Protect Preserved Audio