Key Takeaways
- I was paying a credit every time I generated a voiceover in my own voice. So I asked Claude one question: is there a way to do this for free?
- It built a free AI voice clone that runs on my own GPU, overnight, while I slept.
- The first three takes sounded like me and did not sound like me. Five tuning takes later, one did.
- The hard part was not the voice model. It was getting an AMD card to run it on Windows.
- Once I picked the take, it installed that voice into every video pipeline I have. One command, no credits.
I use a cloned version of my voice on a lot of what I publish. News videos, the faceless channel, the opens on my long form videos. Every one of those cost a credit on somebody else's platform, Higgsfield, ElevenLabs, I honestly lost track of who I was paying. For my own voice. I run On Point Tech Solutions in Los Gatos, and the whole point of my business is that clients own what we build, so paying rent on my own voice started to bother me. This is how I stopped.
One Question, Twelve Hours
The prompt was not clever. I told Claude, running on Fable 5.1, that I already have my voice, I hate paying to generate it, and asked what I would need to build to have it for free. Then I let it run. It came back twelve hours later with a working build. The one decision it made that mattered: run the model on my GPU, not the CPU. It said the CPU path would take about three times longer, and it was right.
The A/B Test Nobody Wants to Sit Through
It rendered the same paragraph in every version so I could compare them honestly. First the original paid clone, then three local takes, A1, A2 and A3. A3 was the closest. A1 sounded like me and did not sound like me at the same time, which is the most accurate review of a voice clone I can give you. A2 was close, no cigar.
While I was listening, I did something I recommend for any tool you are about to build. I found a paid product doing the same thing, Voice Studio, took a screenshot, and fed it to Claude to compare. The answer was that there was nothing it could do my system could not already do. A couple of extra features, nothing we needed. That saved me a subscription I would have bought out of doubt.
The Hard Part Was Windows
The voice model was not the problem. Getting my RX 7900 XTX to run it on Windows was the hardest part of the whole day, and Claude spent a lot of tokens figuring that out so I did not have to. If you are on an AMD card, expect that to be where the time goes. If you are on Nvidia, you will probably skip this paragraph and be done sooner.
Five Takes, One Knob Each
After the A/B, Claude tuned the closest take instead of starting over. Five new samples, T1 through T5, each with one setting turned up or down. Temperature, pitch, a couple of others. T2 was too fast. T3 I did not like at all. T4 got weird at the end. T1 sounded like me all the way through, so I told it T1 it is, lock in those settings.
That was the whole approval. It transcribed the winner, installed it, took it out of the paid pipeline, and started using it in the news videos and the faceless YouTube work the same day. I did not have to do anything else. There is a program sitting on my computer now that generates my voice, and it will be there every time I need it.
Why This Matters for Bay Area and San Jose Businesses
Most small businesses in San Jose that make content hit the same wall I did. The owner is not a voiceover person. I am not either. Sometimes I am just not that good at it, and I would rather have the AI say it right the first time than record it, cut it up and record it again. A free AI voice clone that runs on the machine in your office means the reminder video, the explainer for your website, the weekly update for your clients, all of it can carry your voice without you sitting at a microphone or paying per line.
The bigger lesson is the one I keep teaching at the Academy. When you know what you want to build and you have a blueprint, Claude will work that blueprint like nothing else. The blueprint here was one sentence long.
Practical Steps
- Record a clean reference of your own voice. A minute of you talking normally is plenty.
- Ask for the build in plain English: I have my voice, I do not want to pay to generate it, what do I need. Tell it what GPU you have.
- Insist on an A/B against whatever you pay for now, same paragraph, every take.
- Tune the closest take one knob at a time, and pick by ear. Not the fastest one, the one that sounds like you the whole way through.
- Once you pick, have it install that voice everywhere you generate audio, so the decision is made once.
Watch the Build
You can make your own free voice generation tool with AI. It is crazy, and it took one prompt and a night. If you want systems like this built for your business, On Point Tech Solutions is a Go High Level consultant in Los Gatos serving San Jose and the Bay Area. You own what we build. Done for you, or taught with an SOP. Start at optechsol.llc.