The idea was so clean it should have worked. I had a HeyGen avatar that already looked like me — close enough that my oldest stopped in the doorway, stared at the screen, and said "why are you talking about yourself in third person?" I had a voice clone that sounded like me on a good day with decent mic placement. I had a thirty-second jingle concept for a Roblox promo reel, the kind of thing that lives in a kid's head for a week and costs nothing to produce if you already have the tools. The only thing standing between me and a mildly viral piece of content was four bars of melody.

That was the theory. It was 11 p.m. on a Wednesday. I had a cold cup of coffee and a level of confidence that I can only explain as a side effect of being genuinely tired.

What I got out the other end was not a jingle. It was something between a hostage recording and a malfunctioning GPS. My cloned voice, fed a melody, produced a sound I can only describe as aggressively almost musical. The rhythm was there. The pitch was in the same county. The vibe was a smoke detector that had learned to feel things.

How a Roblox Promo Reel Turned Into a Voice Cloning Experiment at Midnight

I'd been messing with HeyGen for a few weeks at that point — long enough to know its particular flavor of almost-right. I'd already run into the aspect ratio situation that wasted a Tuesday, and I'd spent a full week trying to crank out thirty-second reels with an AI avatar before I understood where the ceiling was. The ceiling is not where they imply it is in the demo videos. The ceiling is lower and has a weird texture.

But the voice clone had been sitting in my account looking useful. For narration, for explainers, for talking-head stuff where I don't want to be on camera at 9 p.m. looking like I've been staring at a compiler error for three hours — it's actually fine. Sounds like me. A little flatter than me, a little too measured, but close enough that I'd used it twice for short content without anyone commenting on the uncanny valley.

The Roblox angle was the kids' idea, loosely. My youngest had been watching these game promo videos on YouTube that all have some kind of earworm in them — a little hook, ten seconds, something that plays over a montage of in-game footage. I figured if I could produce one cheap and fast, maybe it moves some eyes toward the game page. The math on the Roblox hustle has never been good, but we keep adding to it for reasons that have nothing to do with math. So I opened HeyGen, pulled up the voice clone settings, and started poking around for a way to make the thing sing.

There is no "make the thing sing" button. I want to be clear about that. What I found instead was a text-to-speech input and a lot of optimism on my part.

What HeyGen Voice Cloning Actually Does When You Feed It a Melody

The first attempt: I typed the lyrics in with punctuation hacked to suggest rhythm. Commas for breath, hyphens for held notes, all caps for emphasis. The output sounded like me reading a grocery list with unusual conviction. No melody. Just stressed syllables in the wrong places.

Second attempt: I found a workaround someone had posted on a forum — record yourself humming or singing the actual melody, feed that as the voice sample, then use the cloned voice to narrate over it. This is not voice cloning singing. This is layering. I did it anyway. The result was two versions of my voice that disagreed with each other about the tempo, playing simultaneously, like a Jeff argument in a bathroom with good reverb.

Third attempt: I tried a different tool entirely — one of those AI singing generators — and fed it my voice clone's audio file as a style reference. I'm genuinely not sure if that's within anyone's terms of service. I do not know and I moved on quickly. The output was more musical, technically, but it no longer sounded like me. It sounded like a man who had studied my voice from a distance and was doing an impression at a work party.

"Dad, why does it sound like you're crying?" — my oldest, listening to attempt three over my shoulder at approximately 11:47 p.m.

She was not wrong. There was an emotional quality to that third output that I had not intended. The AI singing voice had decided the jingle was a ballad. Four bars about a Roblox game, delivered with the weight of a man saying goodbye.

I rendered it anyway, dropped it into the reel, watched it back twice, and then did not post it. I sat with it for a minute. It was genuinely funny. It was also genuinely unusable. Those two things can both be true, and with AI content they often are — which is a point worth making, because the part that gets skipped in every side hustle video is exactly this: the hour you spend producing something you immediately know you cannot use.

The Real Tab: Three Hours, One Subscription, and a Reel That Never Went Up

HeyGen runs on credits. The voice cloning feature is part of a paid tier. I'm not going to state what it costs because their pricing has changed at least twice since I signed up and anything I write here will be wrong within a month. What I can say is that the three attempts I rendered ate into my monthly credit balance in a way that was noticeable, for output I could not use.

Time: roughly three hours across the evening. One hour of actual experimentation, one hour of me going down a forum rabbit hole about AI singing tools, and one hour of me watching the sad Roblox ballad on repeat trying to figure out if it was secretly good. It was not secretly good. I was just tired enough to find it compelling.

The Roblox promo reel did not go up that week. The game page didn't get the bump. I don't know if it would have helped anyway — the whole operation sits at almost-no-revenue on a good day, and a jingle was not going to change the slope of that line. But I'd talked myself into believing the jingle was the missing piece, and that is a particular kind of wrong that costs more than credits. That's the part where you've convinced yourself the next thing is the thing, and then the next thing turns out to be a guy in a tuxedo crying about your kids' Roblox map.

What I Actually Kept Using After the Singing Experiment Failed

The voice clone, for narration, is still in my rotation. That part works. I use it for short explainer clips where I don't want to sit in front of a camera, and for the occasional reel where I need a voiceover at midnight and don't want to wake anyone up recording in the kitchen. It sounds like me on a day when I've had enough sleep, which is a version of me I rarely have access to in real time.

The singing thing I let go. I looked briefly at a couple of dedicated AI music tools that claim to do voice-style singing from a sample, and they might be worth revisiting someday, but they're a separate subscription for a feature I'd use twice. I've already got one subscription I'm getting mixed value from. The screenshot of the one week it worked great is not a business plan. I know this now. I knew it before, actually. I just needed to make the sad Roblox ballad first.

The jingle idea itself I handed back to the kids. My oldest said she'd just record herself singing it on her tablet, which took her about four minutes and sounded approximately eight hundred percent better than anything I produced. I spent three hours and a non-trivial number of AI credits to arrive at "ask your kid." That's the whole story, honestly. That's the ledger entry.

The reel eventually went up with her voice on it, no AI involved, and it did fine — not great, not viral, fine. A few more eyes on the game page. The sad ballad render is still in my downloads folder. I keep meaning to delete it and keep not deleting it. I don't know what that means.

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.