Voice AI that runs on your own servers.
On-premise voice AI means the whole voice agent runs inside your infrastructure, so your callers’ audio never has to leave it. Lokutor’s models are small enough to run on ordinary servers, with the language model of your choice.
The same voice agent, inside your perimeter.
An on-premise voice AI deployment runs every stage of a call (noise removal, turn-taking, speech recognition and the voice) on servers you control, in your data centre or your own cloud account. It is what regulated teams ask for when call audio cannot be sent to a third-party service: healthcare, finance, insurance, government and anyone under strict data-residency rules.
It is possible with Lokutor because the models are small. Versa, our voice, and our recognition ran live calls from CPU-only servers until September 2026; our cloud now serves them on GPUs because one GPU server carries more simultaneous calls, but neither stage needs one.
- Noise removalPsst, our noise suppression model, cleans the audio before anything listens. Noise cancellation
- Turn-takingTurno decides when the caller has finished and when they are interrupting, on the audio itself: 18,651 parameters.
- Speech recognitionConv: today NVIDIA’s open Parakeet model run on our servers, with our own Conv 2.0 as the next release. Speech to text
- The voiceVersa speaks nine languages in ten voices, or yours cloned from a few seconds of audio. Text to speech
- OrchestrationAn open-source Go orchestrator ties the stages together with barge-in and a pluggable language model: yours, hosted by you or reached through a private endpoint.
- DeliveryScoped and quoted with you: volumes, languages, hardware and compliance requirements, with support terms to match.
Last updated 10 October 2026.
Cloud, on-premise or on-device.
| Our cloud | On-premise | On-device | |
|---|---|---|---|
| Where audio is processed | Lokutor’s servers in the EU | Your servers | The device |
| Setup | Minutes: sign up and build an agent | Scoped and deployed with you | Licensed and ported to your hardware |
| Price | About 5.2¢ a minute, pay as you go | Quoted per deployment | Licence per product |
| Works offline | No | Inside your network | Yes |
| Good for | Most teams, to start | Regulated and high-volume voice | Robots, vehicles, appliances |
On-device is covered on the on-device page; the cloud price is on the pricing page.
Common questions.
What is on-premise voice AI?
Voice AI that runs on your own servers instead of a vendor’s cloud. The full pipeline (noise removal, turn-taking, speech recognition and speech synthesis) runs inside your infrastructure, so call audio does not have to leave it.
Can voice AI run on-premise?
Yes. Lokutor’s models are small enough to run in real time on ordinary servers, so the whole voice agent can run inside your own infrastructure. On-premise deployments are scoped and quoted with you through the enterprise page.
Which voice AI platforms offer on-premise deployment?
Lokutor does. Some other platforms offer self-hosting or a private cloud on request through their sales teams, and others run only in their own cloud, so ask each vendor before you shortlist. Our comparison pages record what each platform says about self-hosting, with the date it was read.
Does on-premise voice AI need GPUs?
Not with Lokutor’s models: speech recognition and synthesis served live calls from CPU-only servers until September 2026. A GPU lets one machine carry more simultaneous calls, so whether you want one depends on your call volume.
How much hardware does it need?
It depends on your volume, languages and hardware, and we size it with you. One reference point from September 2026, for speech synthesis alone: an 8-vCPU server carried three simultaneous synthesis streams at conversational latency, and one NVIDIA T4 carried 36. That counts streams, not calls: a call needs a synthesis stream only while the agent is speaking.
Does on-premise help with GDPR?
Keeping call audio on your own infrastructure means it is not transferred to a processor, which simplifies a GDPR assessment; it does not replace one. Lokutor’s own cloud also runs in the EU: calls are processed in AWS Ireland and accounts and conversation records are stored in Frankfurt.
How much does on-premise voice AI cost?
It is quoted per deployment. For comparison, the same voice agent in Lokutor’s cloud is about 5.2¢ a minute at a typical prompt, paid as you go.
What is the difference between on-premise and on-device voice AI?
On-premise runs the voice agent on servers you operate. On-device runs it inside a product such as a robot, a vehicle or an appliance, offline, on the processor already there. Both use the same small models.
Tell us where it has to run.
Your volumes, languages and requirements. We’ll tell you what fits.