Voice Cloning for Bedtime Stories: How It Works (and Is It Safe?)

A reader emailed me this week with a sentence I keep thinking about: “I want to use voice cloning for my daughter, but the words ‘voice cloning’ make me nervous, and I don’t fully understand what’s happening when I tap that button.”

That captures it for a lot of parents. The technology is fascinating. The technology also sits next to a pile of scary headlines about deepfakes and scam calls. Both things are true, and they’re not separable in most parents’ minds — you can’t just say “this is fine” and have it land.

So this post is for the parents who want to understand. I’m Robin, I built Gramms, and Gramms uses voice cloning. I’ll explain what’s actually happening inside the app, what the legitimate concerns are, what the technology cannot do, and where I think the line is between use and misuse. I’ll be straight about what we do well and where the limits are.

What voice cloning actually does — in plain English

When you record 30 seconds of yourself speaking, you’re not handing the app a giant recording it stitches together. You’re handing it a sample that a model uses to learn the distinctive shape of your voice.

Every voice has a fingerprint. Pitch range (where the voice sits — deep, mid, high). Timbre (the texture — gravelly, smooth, breathy). Pace (how fast you naturally speak). Accent and pronunciation patterns (how you say “water” vs how I say “water”). Cadence (where you pause, how you stress words). A voice-cloning model listens to your 30-second sample and extracts a compact mathematical description of those characteristics.

After that, when the app needs to generate a story, it doesn’t dig back into your recording. It feeds the story’s text to the model along with your voice fingerprint, and the model synthesises new audio that has the shape of your voice but the words of the new story. The story has never been said by you in physical reality. It has been generated from text, in the texture of your voice.

This is why it works for any story length, with any words, instantly. The model isn’t searching a library of phrases you said. It’s drawing your voice the way a pianist who knows your hand could play any new song in your style.

How Gramms specifically does it

Under the hood, Gramms uses Cartesia Sonic for voice cloning. Cartesia is one of the more focused speech-AI labs working on this — audio-only, no video or face cloning, which I want to be explicit about because parents sometimes assume “AI voice” implies “AI deepfake,” and these are genuinely different categories.

The flow inside the app is small and contained:

1. You tap “record voice” inside Gramms on your phone.
2. The app records 30 to 45 seconds of you speaking naturally — any sentence is fine.
3. That sample is sent to Cartesia, which returns a voice profile (essentially a small numerical signature of your voice).
4. The profile is stored encrypted, attached to your account.
5. From then on, when you generate a story, the story text and your profile are sent together, and audio comes back in your voice.

A few specific design choices that matter for safety:

• Phone-recorded only. You can’t upload a random audio file from elsewhere to clone someone else’s voice. The recording has to happen in the app, on the device.
• One account, one set of profiles. Voice profiles you create are bound to your account. They are not browseable by other users, not exposed via any public URL, not exported as raw clip files.
• Used only for your stories. Your voice profile is never used to generate audio for someone else’s story or someone else’s account. It exists in your account to read stories to your own kids.
• Deletable. You can wipe a voice profile (and the original recording) any time from the app’s voice settings.

That’s the whole architecture for the part of voice cloning Gramms touches. It’s deliberately narrow.

Why use a family member’s voice over a professional narrator

A reasonable question: if there are perfectly good professional narrators (and many bedtime story apps use them), why bother cloning a parent’s voice at all?

Because it’s not a content quality question. It’s an emotional anchoring question.

Kids fall asleep to the voice they trust. The voice their nervous system has associated with safety since before they had words. That’s almost always a parent or primary caregiver, and after that, often a grandparent. A professionally narrated story from a stranger is fine — kids can enjoy it the way they enjoy a Disney audiobook. But it doesn’t do the thing a parent’s voice does, which is quietly tell the kid’s body: this is okay, you can rest now.

The use case where this matters most is when the parent isn’t physically there. Working a late shift. Travelling for work. Deployed. Living in a different country from a grandchild. The whole reason grandma’s voice telling a bedtime story hits the way it does is that it stitches connection across distance. The story is just the surface; the voice is the substance.

I get emails from families who’ve used this for a parent doing 70-hour weeks at a hospital, for grandparents in long-distance setups, for military families during deployment. The pattern is the same every time: the kid attaches to the voice, not the production quality.

What concerns parents reasonably have

These are the four I hear most. I’m going to answer them as plainly as I can.

“What if someone clones my voice without my consent?”

This is a real societal concern, but it’s mostly separate from a bedtime story app. The general risk — anyone with 30 seconds of public audio of you (a podcast, a YouTube video, a voicemail) could in principle run that through a cloning tool somewhere — exists regardless of whether you ever use Gramms. Using Gramms doesn’t make that easier or harder; we never publish your voice anywhere.

What Gramms specifically can and does control: the recording you make inside the app stays inside your account. We don’t expose it. We don’t pool it. We don’t use it to train general voice models.

“Is the voice clone available for misuse?”

In Gramms, no. Your voice profile cannot be used by another account, isn’t browseable, isn’t exportable, and is deleted when you delete it. The threat model “someone uses my Gramms voice profile to scam my mother” doesn’t have an attack path that goes through the app — it would require breaching your account or our backend, which we treat as security-sensitive.

“Is the cloned voice indistinguishable from the real me?”

No, and I want to be honest about this. From 30 seconds of audio, the clone captures the things you’d recognise — pitch, accent, general timbre, pace — but loses some of the things that make you specifically you when you’re being expressive. Your laugh. The exact way your voice cracks when you’re tired. Big emotional swings. The cloned voice tends to read a bit calmer, smoother, and more “narrator-y” than your real conversational voice. People who know you well notice. Kids almost universally don’t, in part because they’re hearing the voice in the bedtime context where everyone uses a calmer voice anyway.

“Will my kid be confused or disturbed?”

I worried about this before launching too. In four years of feedback from families using voice-cloned stories, I cannot remember a single report of a kid being upset by it. The closest was an 8-year-old who said “Mum’s voice is a bit different in this story” and the mum explained, “Yes, the app makes a special copy of my voice so I can read to you when I’m at work.” The kid said “cool” and went back to the story. Children handle this much more gracefully than adults expect — possibly because they already accept “Mummy on the phone” and “Mummy on a video call” as versions of the real thing.

If you’re worried, the safest pattern is to be open with your child from the start: the app makes a copy of mum’s voice so she can read to you any time. No fiction required.

What voice cloning is NOT

Three things I want to be clear it isn’t, because conflation does a lot of damage:

• It is not a video deepfake. Voice cloning produces audio only. There is no video, no face, no lip-sync of someone they’re not. Gramms’ tech stack contains no video generation at all.
• It is not a tool for impersonating someone. The app architecture only lets you clone your own voice, recorded on your own device, used in your own account. The “impersonate someone using a voice you recorded of them” use case isn’t a feature of Gramms — it’s a hypothetical attack on the technology category.
• It is not a replacement for being there. A cloned-voice story is a beautiful gap-filler when the real person isn’t physically available. It’s not equivalent to a parent reading to a child in person and shouldn’t be used as one. I say this as a founder of an app that benefits from people using it more — please use it for the nights you’re not there, not as the default replacement for the nights you are.

Use cases families actually report

The patterns I see, from the founder seat, are roughly:

• Working parent on late shifts. Doctor, nurse, restaurant manager, factory shift, founder pulling all-nighters — they record their voice, and the kid gets the bedtime read regardless of whether they’re home.
• Travelling parent. Sales, consulting, tour bookings — the voice goes with the kid through 3-4 nights away from the parent.
• Long-distance grandparent. Granny across an ocean records once and the grandchild gets a bedtime story in granny’s voice every night, with new content. This is the use case that drives the most emotional emails I get.
• Military deployment. Pre-deployment recording, then 6-12 months of stories during active deployment. Massively meaningful for both parents and kids.
• Legacy. A grandparent who knows their time is short records ahead. Or families who’ve lost someone use surviving recordings to give grandchildren a voice they can hear. This is sensitive territory and not for me to tell anyone how to feel about — but the families who use it this way describe it as a gift, not an unease.

How to record a good voice clone — practical tips

If you’ve decided to try it, the recording quality matters more than people expect. A few things that consistently improve the result:

• Quiet room. Background noise gets baked into the profile and reduces clarity. A bedroom with the door closed beats a kitchen with a fridge running.
• Normal speaking voice. Don’t put on your “story voice” or your radio voice. Talk the way you talk on the phone to a friend. The model wants your real voice, not your performance voice.
• 30 to 45 seconds of natural speech. Talking about your day, describing what you’re cooking, explaining what you did at work — anything natural. Reading from a book is fine but tends to produce slightly stiffer clones than free speech.
• No background music or TV. The model will pick up on it.
• Avoid extremes. Don’t yell, don’t whisper, don’t put on accents. Middle-register, normal pace.

Detailed walkthroughs for non-techy family members live in the grandparent voice recording guide.

Voice cloning vs text-to-speech

Worth being clear on the difference, because they’re often confused.

Text-to-speech (TTS) is what Siri, Alexa, and most audiobook generators use. A model is trained on huge amounts of voice data and produces a small set of generic synthetic voices. Some are very good now; some still sound robotic. The voice you hear is not anyone you know. There are usually 5 to 50 options.

Voice cloning uses the same underlying speech synthesis, but personalised. Instead of one of 50 generic voices, the output is in your voice (or whoever recorded the sample). The model conditions its synthesis on your voice profile.

The difference for a child at bedtime is enormous. A generic TTS voice — even a really good one — is a stranger reading them a story. A cloned parent voice is their parent reading them a story. The text might be identical, the production quality might be identical, but the emotional weight is entirely different.

What other apps do this

I’ll be honest about the competitive landscape because pretending there’s only one option doesn’t help anyone trust me.

• Sleepytale offers voice cloning, but only on its higher-priced “Pro Plus” tier, which is roughly $17/month at the time of writing.
• Gramms offers voice cloning included in the standard $5.99/month subscription, with no upcharge tier.
• Most other AI bedtime story apps don’t offer voice cloning at all — they use a small set of professional or generic narrator voices.

If voice cloning is important to your family, the choice is mostly between Gramms and Sleepytale on price/included-features, with the broader landscape covering the apps that don’t offer it. I obviously think Gramms is the better deal — but I’d rather you make that call from honest information than be surprised later.

What’s coming next in voice cloning

Speculating, briefly, because I work near this and parents ask:

• Shorter clip lengths. 30 seconds feels long today; in 18 months it’ll likely be 10. Frontier models already work with 5-10 second samples.
• Cross-language synthesis. Record in English, generate stories in Spanish in your voice. Early versions of this exist today; quality will catch up over the next year or two.
• Emotional range. Currently cloned voices are calmer than the original. Models are getting better at conditioning on emotional context — happy passages sounding happier, suspenseful passages sounding more careful.
• Better safety tooling. The good news is that the same labs working on cloning are working on detection — watermarking generated audio, building “is this synthetic?” classifiers, and so on. The arms race is real but it’s not one-sided.

Whatever happens, the principle Gramms operates on doesn’t change: the only voice you can clone in our app is your own, recorded on your phone, used to read stories to your own kids. That narrowness is a feature.

The honest summary

Voice cloning for bedtime stories is a small, contained use of a much bigger technology category. The bigger category has real risks. The bedtime-story slice — your voice, your account, your kids — is one of the most clearly benign uses available, especially when the architecture is built to keep the voice profile inside your account.

It’s also one of the most quietly meaningful pieces of technology I’ve worked on. A grandparent across an ocean, a parent on a 7pm-to-7am hospital shift, a deployed soldier — these people are getting a way to be in their child’s room at bedtime when they physically cannot be. That’s worth doing carefully.

If you’re still weighing whether AI bedtime stories are safe in general, that’s a separate, broader conversation, and a fair one. This post was about voice cloning specifically — what it does, what it doesn’t do, and where the reasonable line is.

When you’re ready to try it, the recording takes 30 seconds and lives in the voice settings inside the Gramms app.