Recognition, translation, speech
The whole chain runs locally: ASR turns your voice into text, MT translates it, TTS speaks it. No part of it needs a network.
On-device is not the fallback
Sokuji ships free on-device recognition, translation and speech synthesis. It is not there for when the network drops — it is a first-class path, and we expect it to become the default one.
What runs on your machine
An engine matched to your hardware
Sokuji detects your GPU and downloads the engine package built for it, then tells you which one it picked and how large it is.
No key, no account, no meter
Nothing to register for and nothing to spend. Your own hardware is the entire cost.
Why we ship it even though cloud models are still better
We would rather say this plainly than oversell it.
where it stands today
A large hosted model is still ahead on accuracy and on latency, and for a high-stakes call you will probably want one. On-device output is good enough for a great many conversations — and it is the only configuration that involves nobody but you.
where we think it goes
Models keep getting smaller, runtimes keep getting better, and the accelerators in ordinary laptops keep getting faster. We do not expect the gap to last, and we are not waiting for it to close before treating local as a first choice. Nobody should have to hand their conversation to a large AI company in order to get AI at all.
What each route means for your data
This is the only one of the three where the answer needs no qualification.
Your own keyKizuna AI accountOn-device
Audio leaves your machineto the vendorto the vendornever
Someone else can hear itthe vendor canthe vendor cannobody
Works with the network offnonoyes
What it coststhe vendor bills youmetered by usfree