What it is
Vox turns text into speech entirely on local hardware.
It takes whatever the system needs to say — output, a nudge, an ambient note — and renders it
as audio using a neural voice model running on the machine itself. No API calls, no per-request
fees, and no audio ever leaves the box. One command, say "…", and a voice comes out.
Most "natural AI voice" setups reach for a cloud TTS API, which buys quality at the cost of latency, a per-word bill, an internet dependency, and a third party hearing everything. Vox takes the opposite bet: own the whole pipeline, keep it offline, and make it fast enough to feel conversational.
The philosophy
How it flows
The pipeline is fire-and-forget: the command hands text to a local server, which runs inference and plays the result through the system audio in a detached process — so playback survives the caller exiting. A small routing layer can send the audio to this machine, a paired device, or both, from one entry point.
Under the hood
Built on a GPT-SoVITS voice model (voice-cloned from a custom reference profile), driven by Python, wired into the machine through systemd and a shell entrypoint, and played over the local audio stack. Because synthesis is non-blocking and detached, the caller never waits on it — the request returns in well under a second while the voice plays out on its own.
Where it fits
Vox is the delivery mechanism for Anima's whispers. When Anima detects something worth speaking, Vox gives it a voice — and together they make a background presence: something watching, waiting, and occasionally saying hey.