I saw a post on X from @albacnt about VoiceStudio, an open-source project that runs entirely on your own computer. It's pitched as a local alternative to ElevenLabs: voice cloning, text-to-speech, dubbing, transcription, even audiobook production. No subscription, no uploading your voice to someone else's server.

I've written a Western novel, and I've wondered what an audiobook version might sound like. So I handed the job to my AI assistant, Grok Bot, with one instruction: get VoiceStudio running on my iMac.

What I want to share isn't really the tool. It's how little I had to do.

What the assistant handled on its own

It read the docs first. Before installing anything, it went through the project's documentation to understand what the setup needed. Then it installed VoiceStudio from source. The server ran on localhost only, so it was reachable from my own computer and nothing was exposed to the internet.

It worked around a broken link without cutting corners. The hosted installer for the desktop app returned a 404. The assistant read through the repo's own install script before running it, then verified the checksum on the download. It also gave me a heads-up: the app isn't signed with an Apple Developer ID, so macOS may warn me the first time I open it.

It generated and delivered a test clip. About four seconds long: "This is a test of VoiceStudio on Andy's iMac. Peace be with you." The first clip took about 150 seconds, mostly loading the model into memory. For a full-length novel, I'll plan my time around it.

It cleaned up after itself. The install had left a duplicate copy of the source code on the drive. The assistant removed it, which brought the iMac back to 17 GB free.

The one real decision

The assistant stopped only once for anything that mattered. Before downloading a voice model, it flagged that the default model, OmniVoice, is licensed for non-commercial use only. If I ever wanted to sell an audiobook of my novel, that would be a problem. It also noted the iMac was low on disk space, so model size mattered too.

I chose VoxCPM2 instead. It's about 5 GB, allows commercial use, and supports voice cloning.

That's the kind of thing I want to be asked about. It's easy to miss on a project page and painful to discover after you've recorded hours of audio.

My only other input was a quick OK to delete the duplicate copy. Two short answers, and the whole thing took under an hour of my attention.

What I took away from it

A good assistant reads before it runs. It read the documentation before installing and the install script before running it. That's a good habit whether a person or an AI is at the keyboard.

Workarounds should be careful, not clever. When the official installer failed, the fix still went through review and a checksum check.

Check model licenses before you plan commercial use. "Open source" doesn't always mean "free for any use."

Ask at real decision points, and only there. I don't want an assistant that asks permission for every little step. I want one that does the routine work and stops when a choice affects licensing, money, or my machine. That's what happened here.

Local AI keeps your voice yours. Running this on my own machine means the recordings and any cloned voice never leave my desk.

Andy Halvorsen, AWH Industries