The AI voice agent that
never leaves your building.
Vocalis hears, thinks and speaks on your own hardware. No cloud, no API key, no audio leaving your network. The same Zensia guide, for the conversations that must stay private.
Some conversations cannot go to the cloud.
Every AI voice agent on the market streams the caller's audio to a model running in someone else's data centre. For most businesses that is fine. For a bank, a hospital, a court or a government office, it is a non-starter.
Regulation says no
Patient records, account details, citizen data. Data-residency rules and sector regulators decide where that audio may go, and the answer is often nowhere.
The network says no
Air-gapped sites, defence facilities and factory floors run without an internet link on purpose. A cloud agent simply does not work there.
The vendor says maybe
Pricing that changes with someone else's model, outages you cannot fix, and a quota that decides how many calls you may take today. Critical lines need better than that.
Three models. One machine. Nothing leaves.
Vocalis runs a complete speech-to-speech loop on a single GPU inside your premises. Scroll through the three stages.
It hears
- A local speech-recognition model turns the caller's voice into text in a fraction of a second.
- Tuned for Indian-accented names and numbers, the words a cloud model most often gets wrong.
- The audio is processed in memory on your machine and never written to anyone else's.
It thinks
- A local language model decides what to say from your knowledge base, your rules and the conversation so far.
- It streams the reply sentence by sentence, so speaking can begin before the answer is finished.
- Your documents are searched on your hardware. Nothing is uploaded to be indexed.
It speaks
- A local audio model voices each sentence as it arrives. The caller hears the first word before the reply is complete.
- Interrupt it mid-sentence and it stops within about sixty milliseconds, then listens.
- One consistent voice across the whole call, generated faster than real time.
Everything Echo does. Deployed where you say.
Vocalis is not a cut-down product. It is the Zensia guide with the model layer swapped for one you own. Only the audio path changes.
bytes of audio, text or transcript sent outside your network. Not reduced. None.
Knowledge that stays inside
Documents, policies and FAQs are indexed and searched on your hardware. The guide answers from them on every call.
Memory of every caller
It remembers who called and what was said, stored in your database under your retention rules.
Transfer to a human
When a call needs a person, it hands over with a spoken brief, the same way Echo does.
Runs on one machine
The current build needs a single 8 GB consumer GPU. A server card holds several calls at once.
Your transcripts, your logs
Every conversation is recorded and summarised locally, ready for audit without a third party anywhere in the chain.
Teams whose calls carry something sensitive.
Banks and insurers
Account queries, KYC follow-ups and collections calls that must stay inside the regulated perimeter.
Hospitals and clinics
Appointment lines and patient helplines where the conversation itself is a medical record.
Government and public sector
Citizen service desks and departmental helplines with data-residency rules and no tolerance for vendor outages.
Defence and industry
Facilities and plants that run air-gapped by design, where an internet-dependent agent is not an option.
In the lab today.
On your premises with early partners next.
Vocalis runs end to end on a single consumer GPU, in English, with one voice. We measure latency, interruption handling and voice stability every week, and we are opening it to a small number of design partners who need a private deployment and will shape what ships. Hindi and other Indian languages are on the roadmap.
Become a design partner →Vocalis, common questions
Does any audio or text leave our network?
No. Speech recognition, the language model and the voice all run on hardware you control. There is no cloud call in the conversation loop and no API key to configure.
What hardware does Vocalis need?
The current build runs end to end on a single consumer GPU with 8 GB of memory. A server-class GPU lets one machine hold several conversations at once. We size the deployment with you.
Is it the same product as Echo?
It is the same Zensia guide with the model layer replaced by one that runs locally. Agents, knowledge, memory, transcripts and handover to a human work the same way; only the audio path changes.
Which languages does it speak?
English today, with one built-in voice. Hindi and other Indian languages are on the roadmap, and design partners help us decide the order.
Can it connect to our phone lines?
Yes. Vocalis is designed to sit behind the same telephony and PBX connections Zensia uses for Echo, inside your network. We confirm the exact setup during onboarding.
How do we get started?
Vocalis is in early access with a small number of design partners. Leave your business email below and we will reach out to understand your deployment.
Be first to run Vocalis
on your own hardware.
Join the design-partner list. We will reach out personally to understand your deployment and set it up with you.
Business email only · No spam, ever