Deploy it on your own infrastructure.
The whole engine — recognition, translation and voice — runs on your hardware. The audio does not leave your network, and there is no external API to trust, because there is none to call.
What sovereign means here
It is a word everybody uses for something slightly different, so here is ours, in three things you can check rather than three adjectives.
Nothing leaves
Speech recognition, neural translation and voice synthesis all run inside the container, on your host. There is no call out to anybody's API, which is why it also works on a network that has no way out at all.
It is yours to run
It ships as a Docker image with hardware acceleration enabled on the host. Configuration is environment variables and YAML, so it drops into the CI/CD you already have rather than asking you to build a new one.
Compliance by design
When the audio never crosses your perimeter, the data protection question stops being a clause in a contract and becomes a fact about the network. That is a different conversation with your legal team.
Built for Infrastructure
Powered by a proprietary orchestrator engineered for ultra-low latency and massive concurrency. Every component runs on-premise — from speech recognition to neural translation to voice synthesis.
Native Orchestrator
Custom-built runtime for real-time audio pipeline management.
100% On-Premise
Zero external API calls. All models run locally on your infrastructure.
Massive Concurrency
Handle hundreds of simultaneous audio streams with a native concurrency architecture.
Industrial-Grade Security & Resilience
Designed to never fail. Intelligent Circuit Breakers.
An eight-hour session
the brake has acted
MEMORY
CEILING
CPU
CEILING
RESERVED HEADROOM
FLOOR
THE SESSION GOES ON
Memory Protection
Active monitoring that prevents memory saturation, with an operational ceiling that keeps the system stable however long a session runs.
CPU Load Management
Dynamic load balancing across threads with automatic optimization to keep the system always responsive.
Hardware Isolation
Memory held with margin to spare at peak load, so an unexpected burst of traffic can never exhaust it.
And when there is no machine room.
The Liora Box is how you get sound into Liora when there is no rack and nobody to set one up: an event, a conference centre, a mobile unit. A ready-configured physical unit that takes the audio from your desk and sends it to be translated.
- It arrives assembled. Connect the audio and it works — there is nothing to configure.
- This is not the deployment on your premises: the Box captures and sends, and the engine translates on our servers. If what you need is for the audio not to leave your network, what you want is the licence above, on your own hardware.
- Sale of the unit plus the service, and the stations are delivered ready to work.
Technical FAQ
This section addresses the integration, security and performance challenges of Liora Flow.
Start Your Project with Liora Flow
Tell us about your technical challenge. Our engineering team will analyze your use case to propose the most efficient inference architecture.
The jargon of your trade
The engine hears everyday language well. It fails on the vocabulary used only inside a profession — «paroxysmal nocturnal dyspnoea», «plea of lis pendens», «tightening torque» — and it always fails the same way: it swaps them for ordinary words that sound similar.
- It is tuned with material from a field, not from one house: paroxysmal dyspnoea is the same in every hospital. That is why shared tuning is not the cheap option, it is the one that works better.
- The glossary comes first, and it is almost always enough. This is the step above, for whoever already tried it and came up short.
- It is not switched on with a checkbox: you talk, you get a quote, you sign. And it switches off with a button, no questions, immediately.
On your own servers you need a card with at least 24 GB: the model does not fit on a desktop one.