I Cloned My Voice in HeyGen and Kept Talking

It was a Tuesday night, kids in bed, laptop on the kitchen table, and I was recording myself saying the word "banana" for the fourth time into a USB mic I bought two years ago for a podcast I made three episodes of. The HeyGen interface was asking me to read a calibration script out loud — slowly, clearly, no background noise — and my ceiling fan was doing its best to ruin everything. I paused the fan. Read the script. Submitted the sample. Waited.

The voice that came back sounded like me after a good night's sleep, which is a version of me I have not personally met in some time. It had my cadence, my slight Southern-California-that-moved-to-Alabama flatness, even a little of the way I trail off at the end of a sentence like I'm not totally sure I'm right. Which, to be fair, I usually am not.

I had been messing with HeyGen on and off for a few months by that point — long enough to know the aspect ratio situation can quietly ruin a batch of videos before you notice, and long enough to have real opinions about the loading spinner. The voice cloning feature had been sitting in my queue the whole time. I kept putting it off because it felt like the kind of thing that either works immediately and changes the game, or works immediately and makes you feel weird about yourself. Both of those turned out to be true.

Why I Bothered Cloning My Own Voice Instead of Just Talking

The honest answer is I was already making short-form video content — mostly Reels filmed in the car, which is the only place I can talk to a camera without someone walking through the background — and I kept running into the same problem: I didn't have time to record a new voiceover every time I had a script idea. I'd get an idea at 11pm, write it down, and by the time I had fifteen minutes to actually record it, the energy was gone. The idea sat in a Notepad file from 2023 and waited.

Voice cloning felt like a fix for that. Write the script, paste it in, let the clone talk. I'd done a whole week of letting HeyGen handle my Reels pipeline and the bottleneck was never the avatar — it was always me, either recording audio or approving takes or just not being awake enough to care. If I could hand the voice part to a trained copy of myself, I could theoretically ship more. That was the theory.

My brother-in-law had already texted me about it. He'd seen something online and wanted to know if we could "build a channel that just runs itself." I told him I'd look into it. He sent three more texts before I finished reading the first one.

"Bro you could literally post while you sleep."

I did not tell him that was the part I was skeptical of. I just said I'd report back.

The Clone Worked. That Was the Whole Problem.

HeyGen voice cloning — the actual technical experience of it — is genuinely impressive in a way that made me sit back in my chair and stare at the ceiling for a second. You feed it a clean audio sample, it processes, and then you type words and it says them in something that is recognizably your voice. Not a robot impression of your voice. Not a text-to-speech voice that went to the same school as you. Your voice, approximately, doing the thing you told it to do.

I typed in a sentence I had never actually said out loud: "The Auburn offense looked sharp in the second quarter." The clone said it. It sounded like me saying it. I played it back three times. Then I played it back a fourth time while eating leftover Chinese takeout because I didn't know what else to do with my hands.

I built out a few test videos. Avatar on screen, cloned voice reading the script, decent enough B-roll. The results were fine. Not great — the lip sync had the usual slight uncanny-valley shimmer that HeyGen has when you look too closely — but for a phone screen at normal scroll speed, they passed. I posted two of them. One got a reasonable amount of views for me, which is not a number I'll brag about. The other one got twelve.

The thing I kept running into was a different problem than I expected. It wasn't the quality. It was that I still had to write the scripts. I still had to edit the output. I still had to decide what the video was actually about. The voice clone removed one step out of maybe eight, and that one step was not the slow part. The slow part is having something worth saying. A clone of my voice reading nothing useful is still nothing useful, just delivered more efficiently.

I also tried to see if the cloned voice could handle anything musical — I'd been curious about that ever since I went down a rabbit hole on whether HeyGen voice cloning can actually sing. Short answer from my own experiment: no. The clone speaks. It does not perform. Ask it to carry a melody and it gets very polite and very wrong at the same time.

What I Actually Spent on a Voice That Sounds Like Me

I'm not going to pretend I tracked this perfectly. I don't know what I should tell you about HeyGen's pricing because their tiers move around and I don't want to state something as fact that's already outdated by the time you read it. What I can tell you is that I was on a paid tier when I did this, and the voice cloning feature was part of it, and the cost felt reasonable right up until I asked myself what I was actually getting back out.

The USB mic was already paid for — legacy cost from the podcast era, which I'm counting as a sunk loss and not a current expense. I spent probably three hours on the calibration, the test videos, the posting, and the staring-at-the-ceiling portion of the evening. I made nothing directly from those two videos in any trackable way. The one that got views got views. Nobody bought anything. Nobody signed up for anything. There was no anything to sign up for, which is maybe the more honest version of the problem.

Three hours is a long time when you're also a remote software engineer with a standup at eight the next morning. I have a whole ledger of moments like this — time that didn't convert, effort that just sort of evaporated — and this one goes in the same column as all the others. The hours I don't get paid for have their own quiet running total that I try not to add up too carefully.

What I Kept and What I Let the Clone Keep

I still have the voice model. I use it occasionally, when I have a script that's already tight and I just need it delivered without me having to sit down with a mic at midnight. That specific use case — script is done, I just need audio — is genuinely useful. It's not a shortcut to having ideas. It's a shortcut to recording ideas I already have, which is a narrower thing than it sounds.

What I dropped was the idea that this was going to change the throughput of the whole operation. The brother-in-law wanted a channel that runs itself. That's not what this is. What this is, is one fewer step in a process that has a lot of steps, most of which still require me to be awake and thinking.

I also dropped the guilt about the weird feeling. The clone sounds like me. I've made peace with that. It's a tool. The ceiling fan is also a tool and I don't feel strange about it. I'm drawing a line somewhere around "ceiling fans don't say things I didn't write," but I wrote the scripts, so I'm counting it as close enough.

What I didn't figure out — and still haven't — is the monetization side of any of this. I have views. I have a voice that sounds like me. I do not have a clear path from those two things to a number that would make the three hours feel justified. That might be a me problem. It might be a content problem. It might be that I keep trying to build the output engine before I've figured out what the output is actually for. I know better than to buy the gear before the first dollar shows up, and somehow I keep finding new ways to do it anyway.

The clone is still in my HeyGen account, waiting. I have four half-written scripts in a Notepad file. The ceiling fan is back on. At some point I'll finish one of the scripts and let the clone deliver it, and maybe that one will be the twelve-views one, or maybe it'll do something. I genuinely don't know, and that's where I'm leaving it.

Note: Everyday Hustle Jeff is the written side of a digital series. These accounts are composite and dramatized. Any figures are what they were on the day, for one person, and are not a projection, a track record, or advice.