The first time I heard “another me” speak, I’ll be honest—it was unsettling.
That was three minutes after I fed my voice into a cloning tool. It read a script I had never written, in my cadence, with my habits. Ninety percent me. The missing ten percent was hard to name—something flat around the edges. Still, it was enough to make me decide on the spot: I’m testing this properly.
Over the next two weeks I produced thirty voiceovers with my clone—short-video narration, course material, product explainers, an audiobook-style long read, plus a few of my own terrible jokes. This post is the full report: where it works, where it doesn’t, and the sample-recording traps nobody warned me about.
Why I wanted my own voice cloned
The reason is boring: output volume.
I publish four to five short videos every week, and I record all the narration myself. The problem is that human voices aren’t machines. One late night, a cold, dry air—the next day’s recording is unusable. Worst case ever: I re-recorded a two-minute video five times because each take sounded like a different person.
Swapping in a stock AI voice fixes the recording problem but breaks something subtler. Subscribers follow you, not a pleasant stranger. So voice cloning isn’t a party trick for me; it’s turning my voice into an asset I can back up. On bad voice days, the channel still ships—and it still sounds like me.
Recording the sample
The process is simpler than I expected: read the prompted sentences into a microphone, about ninety seconds, upload, wait twenty minutes.
Simple does not mean casual. My first sample was recorded at my desk, and the clone came back with a weird rattle on certain consonants. It took me a while to find the cause: the air conditioner was blowing straight into the microphone. For take two I did three things—turned the AC off, kept the mic one fist away from my mouth, and recorded in the quieter part of the afternoon. Night and day.
The other factor is variety. If your sample is all flat statements, your clone will only ever speak in flat statements. On the second take I kept my natural ups and downs in—questions, a little laughter—and the result came back noticeably alive.
Where the clone earns its keep
After thirty voiceovers, the winner is clear: informational content.
Short-video narration is my top use. A three-minute script takes about five minutes total now, and friends can’t tell the difference. Course narration is the other natural fit—twenty-minute stretches of dense, low-drama material where the clone is actually more consistent than I am; I tend to get tired and pitchy by the back half. Long reads work too. Listening to my own article as audio during a commute is strange, and genuinely useful.
The hidden win is revision cost. Before, a one-word client edit meant re-recording the take and matching the tone of the previous line. Now I generate the same sentence three times and pick the best one. If you have perfectionist tendencies, this is worth more than you think.
Where I still record myself
Ads, first. Clients pay for human warmth—the shifts between excitement, restraint, and pause. Current clones give you one emotion, all the way through. Fine for tutorials, wrong for a brand film.
Second, anything live: streams, interviews. Cloning presumes a script. No script, no show.
Third, a surprise: very short lines. Two-second taglines come out of the clone without the whip-crack at the end. It handles long sentences beautifully and fumbles short ones. My division of labor is now fixed: the clone carries the body of the content; I record the openers, the closers, and the one line that has to land.
How much the sample matters
More than anything else in this post.
Same tool, same script: a casual sample versus a careful one sounds like phone speaker versus studio monitors. Noise in your sample gets learned into your voice. So does your reading-out-loud voice—if you record like you’re reading a hostage statement, your clone reads like one forever.
Habits that survived testing: warm up by chatting for a few minutes before recording; keep pauses natural instead of measured breaths; and stop using the mic that came free with your headphones—even a modest lavalier or a basic condenser changes everything. None of this is expensive. All of it shows up in the output.
The two questions everyone asks
“Can people tell it’s AI?” For informational content, mostly no. I blind-tested ten friends; two said something felt off, but couldn’t say what. Emotional content is different—that ten percent gap gets loud.
“Can I clone someone else’s voice?” No. Full stop. Cloning another person’s voice requires their consent—platform rules in the best case, the law in many places. This connects directly to voice licensing, which I covered in my explainer on how TTS works. My personal rule: only my own voice, and commercial projects get signed paperwork.
Two weeks in: the time math
Numbers argue better than adjectives. Before the clone, a typical three-minute narration cost me about forty minutes: fifteen recording, ten cutting breaths and flubs, fifteen cleaning noise. Now the same job is twelve: five to generate, two to pick a take, five to fix whatever the winning take didn’t cover.
One honest caveat: “generate” is plural. I always run two or three takes and choose, which means build that into your estimate—the tool stays honest with you as long as your math does.
Then there’s the gain that never shows up on a timesheet: ideas don’t go stale anymore. A script written at eleven p.m. used to wait two days for my voice to cooperate, and by then the energy in it was gone. Now it ships while the thought is still warm. The voice never has an off day either—day thirty sounds exactly like day one, every syllable in the same register. I’ve essentially outsourced the manual labor to yesterday’s version of me, who happened to be well-rested.
Three mistakes worth skipping
Mine, so you can skip them. First, I asked the clone to act. Gave it an emotional script with arcs, builds, and pauses, and got back a competent monotone with feelings painted on. It narrates beautifully; it does not perform. That discovery cost me one wasted afternoon and reshaped my whole division of labor.
Second, early on I shipped the first take of everything. Two extra generations and a pick would have lifted every single one of those videos, for free. Laziness has a price and it’s ten minutes.
Third, my very first sample was recorded on the mic that came with my headphones, and I burned an entire redo cycle finding out what it did to the clone. If you buy one thing before cloning your voice, buy a real microphone. It’s the cheapest quality multiplier in this whole process.
If you want to try it
Three tips. One: take the sample seriously—ninety seconds decides the quality of everything that follows. Two: start with low-drama content (narration, courses, explainers) and work up to harder material once you know the tool’s temper. Three: treat your clone like a new hire—give it simple jobs first, learn its limits, and you’ll naturally know when to step back in yourself.
Soundwaver’s voice cloning works exactly this way if you want to try: record your sample, upload, done.
And one combo tip: if you publish weekly like me, pair the clone with these eight short-video voiceover fixes—one handles your script’s rhythm, the other handles voice supply. Together, they cut the pressure of shipping in half.
