Local Voiceover Lab

Why I tried it

Local Asset Forge started as a question about making game assets on my own hardware. Along the way, I tried local audio as well. The 3D work was interesting, but the voiceover side made me think about a more immediate problem: I want to make more videos, and rerecording narration every time I change a line slows me down.

I wanted to know if I could make a voiceover that sounds enough like me to put in an actual video. It had to be something I could run locally and revise without starting the whole recording over.

My setup

I used the open-source VoiceStudio app with its VoxCPM2 voice model. I work on my MacBook, but the model and backend run on my RTX 4080 laptop in another room. My Mac was short on space for the model downloads when I set this up, so I connected the app to the laptop through a private SSH tunnel. The recording, editing, and review still happen at my desk.

I recorded a natural passage in QuickTime and kept its original file. A short section became the voice reference. I kept a different line from that same recording out of the reference so I could compare my real voice with a generated version saying exactly the same words. The video labels both sides of that test.

The retake

The first version was promising, but I could hear that it wasn’t quite right. A cleaner sample, spoken the way I normally narrate, made a noticeable difference. I generated the new narration in separate sections, listened to each one, and rebuilt the video around the takes I preferred.

At about 0:19 in the film you hear the real recording. At about 0:24 you hear the local clone repeat it. Apart from that real comparison line, the narration in the finished video is generated from my voice sample. The on-screen VoiceStudio interface is a labelled still from the tool’s official demo, rather than footage of that specific generation.

Would I use it?

Yes. I already used it to narrate this video, and I plan to use it again. It gives me a practical way to try another take when I change a script, while keeping the model on hardware I already own.

It still needs a listen. A sentence can have all the right words and still miss the timing, emphasis, or personality I wanted. If that happens, I can generate it again or record the line myself. That’s a much more useful result for me than a perfect-looking demo that I wouldn’t put in a video.