How it works

An incoming call, end to end

From the first ring to the owner hearing about it: every service a call passes through, what each one does, and how they connect.

The whole call at a glance

  1. 1 Caller to Carrier Dials the shield number
  2. 2 Carrier to Voice Signed webhook, then the audio stream
  3. 3 Voice to Data Whose number is this? Rules and memory
  4. 4 Voice to Speech Caller audio
  5. 4 Speech to Voice They finished speaking: the words
  6. 5 Voice to Model Words, allowed actions, memory
  7. 5 Model to Voice A decision, then sentences
  8. 5 Voice to Speech Each checked sentence, to be voiced
  9. 5 Voice to Caller The reply, in the owner's voice
  10. 6 Voice to Owner When the rules say so: ring the owner's own phone
  11. 7 Voice to Data Transcript, recording, a note for next time
  12. 8 API to App Signed event: a call happened
  13. 8 App to Owner The call, the voicemail, the messages

Scroll to see how it flows

Step 1:A call comes in

  • Phone network
  • Caller
  • Carrier

Someone dials the owner’s shield number: a second number that the owner’s double answers.

  • One shield number per person. One person, one account, one number.
  • Rings first. Two rings by default before anything answers, so it feels like a phone.

Step 2:The carrier hands it over

  • Phone carrier
  • Carrier
  • Voice

The phone company delivers the call as a signed request, then opens a live audio stream.

  • The number belongs to a carrier. A call is answered only when it arrives through the carrier that owns that number.
  • Signed webhook. Every request is checked against the carrier’s signature before anything else happens.
  • Live audio. Phone-quality audio flows both ways over one open connection.

Step 3:The voice engine picks up

  • Voice engine
  • Voice
  • Data

The conductor of the call. It owns everything about a conversation except running the AI models.

  • Who is being called. Finds the account behind the number without ever storing the number in the clear.
  • Who is calling. Works out the caller from the contact list and the carrier only, never from what the caller says.
  • Call session. One state machine per call: connecting, greeting, listening, thinking, speaking, interrupted, ending.
  • Greeting. The opening line is voiced ahead of time, so it plays at once.
  • Database. Every row belongs to one account, and the database itself refuses to show it to another.
  • Key vault. Each account has its own key. Destroying it makes that account’s data unreadable everywhere, backups included.

Step 4:Listening

  • Speech service
  • Voice
  • Speech

The speech service hears the caller, notices the moment they finish, and turns their words into text.

  • Knows voices, not people. The speech service never learns which account a call belongs to.
  • Hearing the end of a turn. It tells a finished thought from a pause, and "mm-hm" from a real interruption.
  • Speech to text. The caller’s words arrive as text within a fraction of a second.

Step 5:One turn of the conversation

  • Under a second
  • Voice
  • Model
  • Speech
  • Caller

From the caller’s last word to the first syllable of the reply. The goal is well under a second.

  1. 1 Narrow

    The owner’s rules for this caller cut every possible response down to the few that are allowed.

  2. 2 Think

    The language model reads the words, the allowed actions and what it remembers about this caller, and starts answering while still thinking.

  3. 3 Decide

    The first thing it says is a decision: keep talking, or hand the call somewhere. It can only pick from the allowed list.

  4. 4 Check

    Every sentence is checked before it is spoken: nothing private, nothing the owner banned, nothing made up, and an honest yes if asked "is this an AI?"

  5. 5 Speak

    Each sentence is voiced the moment it is ready, in a voice cloned from the owner with their recorded consent.

Step 6:What the moment needs

  • The owner’s rules
  • Voice
  • Carrier
  • Owner

The owner decides, per contact, what their double may do. The call takes one of four paths.

  1. 1 Keep talking

    Answer, ask, arrange. The turn repeats until the conversation ends.

  2. 2 Take a message

    A greeting in the owner’s voice, then the caller leaves a voicemail.

  3. 3 Ring the owner

    The call is passed to the owner’s own phone, showing the real caller. No answer falls back to voicemail.

  4. 4 Decline

    A polite no, or a busy signal, for callers the owner does not want to reach them.

Step 7:After the call

  • Memory, per contact
  • Voice
  • Model
  • Data

The call is written down, condensed into a short note for next time, and counted.

  • The record. The transcript and timings are stored encrypted under the account’s own key.
  • A note for next time. One short running note per contact, so the next call picks up where this one ended. Nobody else’s carries over.
  • A second look. The call is reviewed for anything the owner should see.
  • Audio. Recordings and voicemails are kept as files, never in the database.
  • Minutes. Usage is counted in minutes, the unit customers think in.
  • File storage. Voice samples, recordings and message media live here, reachable only through the API.

Step 8:The owner finds out

  • Platform API
  • Your app
  • API
  • App
  • Owner

The owner never signs in to Ovance. Their app asks the platform API and shows the call under its own brand.

  • A signed event. The app is told that something happened, with ids and counts only, never content.
  • Read live, never copied. Transcripts, recordings and contacts are read from the API each time, so deleting them at the source deletes them everywhere.
  • Keys stay on the server. The app’s API keys never reach a browser. Audio goes straight to the voice engine with a short-lived token.

The whole system

Everything a call touches is lit. The rest is what stands around it.

Cyan edge below: on the call path

The edge

  • Domain and DNS The domain every service answers at.
  • Mail Mail sent and received at the domain.

Our services

  • Voice engine Calls and the turn loop.
  • Speech service Hearing and speaking; knows voices, not accounts.
  • Platform API What apps built on Ovance call, server to server.
  • Console For Ovance’s own administrators only.

Data

  • Database Isolation per account, enforced by the database itself.
  • Key vault Destroying an account’s key is the deletion receipt.
  • File storage Audio, voice samples and media, never in the database.

Models

  • Speech models The cloned voice and speech-to-text, hosted, called from the speech service.
  • Language models, cloud Each task is bound to the model suited to it.
  • Language models, local A floor on our own hardware, for accounts whose data must not reach a cloud model.

Carriers

  • Phone carrier Numbers and calls.

Beyond the call

  • Onboarding and voice A guided conversation records the owner’s voice with consent and learns how they talk.
  • Texts and replies The same persona answers texts and drafts replies, with one memory per contact.
  • Deeper recall Past conversations can be searched by meaning, under the same isolation.