A three-part build: a voiceover studio that clones your own voice from ten seconds of audio, with open-weight models on one GPU in your own AWS account. Part 1 explains how voice cloning works and gets it running, Part 2 opens up the pipeline and the UI, and Part 3 is the AWS side and the real bill.
Ten seconds of my voice, a description of the video, and a voiceover that sounds like me. How voice cloning actually works, why I didn’t just use an API, and how to run the whole thing in your own AWS account.
What actually happens between “describe your video” and a finished voiceover: a core that doesn’t care which model you use, a script written for the ear, one sentence at a time, two models sharing one GPU, and the UI I got wrong the first time.
What one deploy command actually builds in your AWS account, who is allowed to press it, the four layers that stop a forgotten GPU from becoming a $600 bill, what the security scanners found, and everything the first real test run broke.