I Cloned My Grandma's Voice From a Noisy Kitchen Recording — and It Went Badly

Last Thanksgiving I decided to do something I thought was romantic.

My grandma was 84 that year. She grew up in a hollow in eastern Kentucky and she still talks like it — vowels flattened out, "you'uns" when she means more than one of you, "warsh" for wash. Nobody under fifty in our family talks that way anymore. So I got it in my head to record her, clone the voice, and turn a letter I'd written into an audio keepsake we could play every Christmas.

I grabbed my phone and recorded her for half an hour. In my kitchen. On Thanksgiving. Dishwasher running, oven fan going, the game on in the next room, my nephew shrieking about something.

Then I dropped the file into a voice cloning tool and waited.

What came out... how do I put this. It sounded like my grandma with a head cold, a mouthful of cornbread, and a bad phone connection from the bottom of a well.

Attempt number one: dead.

The tool wasn't the problem. My recording was.

My first instinct was the classic one — blame the software, go find a better one. Browser-based ones, ones with a desktop app, the ones advertising free voice cloning, the ones that want your card before you hear a single syllable. I ran through most of them.

Same tunnel sound every time.

That's when I had to admit the model wasn't broken. The half hour I fed it was. In that file, my grandma's voice was maybe 30 percent of what was happening. The rest was the dishwasher, the oven fan, a dog two streets over, and the bounce of my own kitchen tile. You are what you eat — and the model ate a dishwasher and filed it under "grandma."

Three mistakes, each dumber than the last

I recorded in the worst room in the house. Kitchens are reverb machines: hard tile, glass, stainless steel, sound pinging around. A dialect lives and dies on pitch detail, and reverb smears exactly that. A small bedroom with the curtains pulled and a blanket on the bed beats a kitchen by a mile. A closet full of coats is better still, if you can talk someone into standing in it.

I made her read a script. I printed out the letter and asked her to read it aloud. She went slow and stiff, spitting out one word at a time like a kid doing the announcements at church. That is not how she talks. The run-ons, the swallowed syllables, the way her voice trails off at the end of a sentence — gone. So the clone came back sounding weirdly proper. Proper is the one thing my grandma has never been.

I assumed longer meant better. Half an hour of audio, I figured, is more insurance. In reality maybe eight minutes of it was usable. And because the phone was propped on the counter, it drifted closer and farther the whole time — volume all over the place. That kind of input gets you a vague outline of a person, not a person.

The re-record day

A week later I set up in my sister's spare room, the north-facing one that's always cold and always quiet. Lessons applied:

  • Phone on a tripod, about eight inches from her mouth, and it did not move once
  • Windows shut, AC off, curtains drawn, phone in airplane mode
  • No script. I asked questions instead — "What was Christmas like when you were little?" "How did your mom yell at you when you broke something?"
  • Five to eight minute chunks, listening back immediately, deleting and redoing anything with a hum in it
  • Roughly twenty minutes of clean audio total, across four sittings

While I was at it, I previewed the letter text with a free text to speech online tool first — no cost, no signup, nothing to lose by experimenting. That step mattered more than I expected. Where the "well, I'll be" lands, where the "honey" goes... those little words are the whole soul of the letter. You want them in the right place before you commit.

Fair warning if you go shopping around: text to speech online free is a crowded, slightly shady field. Half of them watermark the output or cap you at 200 characters and then act surprised when you're annoyed.

For the real thing, I uploaded the twenty minutes and waited a few minutes.

The first line that came back was "Hey there, sweet pea." My sister was standing behind me and went completely still. The last vowel dropped the way it always did, with a little rasp under it.

That was her.

What the finished thing actually sounds like

I broke the letter into 26 segments, a sentence or two each, from her childhood through the years she spent raising me. With that many pieces I really wanted unlimited free TTS — nothing kills the mood like getting cut off at segment 19. The site I used runs on no registration text to speech, which meant my sister could open it on her ancient laptop and just... start. No installs, no sketchy "free voice cloning software" download that turns out to be neither free nor software.

One thing I'd tell anyone: stop chasing perfect audio. I kept the breaths. I kept the little mouth clicks. Without them you get a news anchor, not your grandmother. That's the whole difference between clean and realistic AI voice online.

If your relative's accent is thick — and mine is — even clean audio will drift in places. Don't fight it. Pick the five or six segments where she sounds most relaxed and let those carry the training, then fill the rest with a standard-English pass. Family won't notice. Family will cry either way.

And if your family is scattered across languages the way mine is — my cousins married into a Spanish-speaking household — multilingual TTS free is the difference between one keepsake and three.

Doing this at scale? There's a free voice clone API for that: push the text in, pull audio out, and a 40-segment project is done before your coffee gets cold. Same story if you want free AI dubbing for a video, or a free AI voice generator for a channel. For a once-a-year project like mine, clicking around a web page is plenty.

FAQ

I already recorded in a noisy room. Can I save it?

Partly. Run it through the noise reduction in whatever free AI dubbing or editing tool you've got, then cut anything obviously ruined and keep only the clean stretches. But here's my rule: if you end up with less than three minutes of clean audio, re-record. Two hours of re-recording beats two weeks of audio repair that might not work anyway.

Why is a dialect so much harder to clone than standard English?

Data. Models have heard a million hours of broadcast English and almost none of your grandmother's specific corner of the world. So every second you give it has to be pure. Dialects also carry pitch patterns that don't exist in standard English at all, which means the model needs more examples before it stops guessing. Casual conversation gives it far more to work with than reading ever will.

How far can free tools actually take me?

Further than you'd think. A family audio letter, a dialect voiceover for a YouTube video, a TikTok or Reels narration — free speech synthesis handles all of that now, and honestly it's better than I expected going in. Batch work is doable too, if you're willing to wire up an API. What free tools won't do is hold up a 90-minute audiobook with a consistent character voice. That's a paid problem. For the rest of us, online voice cloning free is genuinely enough — and when the cost of trying is zero, you might as well try.

Don't wait on it

My grandma passed that October, about five months after we recorded.

The letter lives on my phone and in the cloud now. When I played it this past Christmas, my mom cried, and my six-year-old niece looked around the room and asked who was talking.

She couldn't understand a word of it. That's fine. When she's older I'll tell her: that's your great-grandma, and that's what home sounds like.

If you've got someone like that still around, don't put it off. Pick a quiet afternoon. Kill the AC, prop up your phone, and ask her what Christmas was like when she was little. The software part is easy — I've got more walkthroughs on the updates page if you want them.