Providers

Where the translation actually happens

Sokuji does not do the recognition or the translation itself — it captures your voice, sends it somewhere, and speaks the answer back into your call. You choose where that somewhere is. There are three routes, and you are never stuck on one.

Three routes

On your own machine

Download an inference engine matched to your GPU and run recognition and translation locally. Free, works with the network off, and nothing leaves the device — not even to us.

free · no account · offline

Your own API key

Sign up with any provider below, paste the key into Sokuji, and pay that provider directly. We wrote a walkthrough with screenshots for each one. Nothing routes through us at all.

you hold the key · you get the bill

Through a Kizuna AI account

For people who would rather not deal with developer consoles. We hold the key and settle the bill; OpenAI, Doubao or Soniox still runs the model. Your audio goes straight to them — we are not in the path.

metered · no subscription

Supported providers

Sokuji ships adapters for these. “Speaks” means it can produce spoken translation rather than text only — which is what the flagship scenario needs. Where a Kizuna AI account is offered, the vendor actually running the model is named, because that is who receives your audio. “Face-to-face” means two people at one microphone, each hearing the other’s translation in one earbud; the provider that can do it also tells several speakers on the other side apart.

ProviderHow it runsSpeaksFace-to-faceSetup
Local inferenceOn your machine, engine downloaded to match your GPUyesnoGuide
OpenAI RealtimeYour key, direct to OpenAIyesnoGuide
OpenAI TranslateYour key, direct to OpenAIyesnoGuide
OpenAI LiveYour key, direct to OpenAIyesnoGuide
Google GeminiYour key, direct to GoogleyesnoGuide
SonioxYour key, or a Kizuna AI account — vendor: SonioxyesyesGuide
Palabra AIYour key, WebRTC transportyesnoGuide
Doubao AST 2.0Your key, direct to DoubaoyesnoGuide

What the account costs

Only this route costs anything. Sokuji itself is free and open source, and on-device inference is free with it.

metered, never monthly

You top up a balance and it is drawn down as you use it. A month you do not use costs you nothing, and there is no seat count.

There is no single price to quote here, because each vendor meters differently — some by audio duration, some by what the model processes. The rate that applies is the one shown against that vendor in your account, before you spend anything.

still not in the path

Paying us does not put us between you and the model. We hand your device a credential and count what it spends; the audio goes device to vendor exactly as it does with your own key. See how that is wired →

Balance is valid for one year. That bound exists because unspent credit sits on our books as a liability, not to push anyone into spending it.

Speak your own language.
Be heard in theirs.

free and open source · on-device ai included · no subscription