Local Voice Inference

Lab · R&D, not a commission

Speech-to-speech on my own machine, with open-weight tools

Year: 2024
Role: Voice pipeline and source recording
Tools: RVC-style voice conversion, Gradio, Open-weight models

Five characters, one recorded performance, converted locally. 1:14, with sound. Made in 2024: local models have come a long way since, mouths especially. Listen to the voices, forgive the lip sync.

01 What it is

Speech-to-speech, run entirely on my own workstation. I recorded the source audio myself and converted it with a custom toolset: an RVC-style voice model setup driven through a Gradio interface. The converted voice changes from character to character; nothing was sent to a cloud service.

02 Why local

Commercial voice services do this well and charge for it. My point in the film is that a capable GPU, free tools and a bit of knowledge get you the same kind of result at home, and that is the knowledge I want to pass on.

Next in Lab · 03 / 07Pixelated Portals