Rubricon Voice

Agents that hold a phone conversation, not a menu tree.

The agent answers or places the call on the telephony you already run, in the language the caller chooses to speak. It listens while it talks, looks things up mid-sentence, and hands to a person the moment the conversation leaves what it is allowed to do.

What it handles

Bounded call types, one at a time.

Every deployment starts with a single call type with a clear end state. An agent that tries to handle everything is an agent nobody can score.

Inbound

Calls that arrive

Order, delivery and ticket status, read from the system of record
First-line support with escalation on anything out of policy
Qualification and routing before a person picks up
Overflow and after-hours, with the same rubric applied

Outbound

Calls that go out

Delivery and appointment confirmation, with reschedule handled on the call
Renewal, service and payment reminders
Post-service feedback captured as structured fields, not free text
Callback on an abandoned queue, dialled when the caller asked for it

How a turn works

The pipeline is streaming end to end, because the pause is what gives it away.

A caller reads a gap after their sentence as the other party not listening. Every stage below emits partial output rather than waiting to finish, so the answer starts forming while the caller is still speaking.

Continuous

Media in

Audio off the SIP leg in 20 ms frames, over a persistent socket.

≤ 200 ms

Endpointing

Voice activity plus a semantic check on the partial text, so a mid-sentence breath is not treated as the end of a turn.

Partial

Recognition

Streaming ASR emitting partial hypotheses; the model sees text before the caller stops.

≤ 400 ms

Model

First token, with retrieval and tool calls issued in parallel rather than in sequence.

≤ 200 ms

Speech out

Streaming TTS on the first clause, so playback begins before the sentence is composed.

Any time

Barge-in

Caller speech cancels playback and discards the queued audio mid-word.

These are the budgets the pipeline is built to, not a benchmark. The measured per-turn distribution for your call type is reported from the telemetry, per deployment.

Controls

What the agent is not allowed to do is the part that gets configured.

Escalation: Warm transfer to a named queue, with the transcript and the retrieved record on the receiving agent's screen before they say hello. Triggered by a topic leaving scope, a compliance rule firing, a caller asking, or two failed attempts at the same step.
Grounding: Answers about policy, pricing and availability are retrieved from your documents and systems at call time. Anything the agent cannot ground, it does not assert — it says so and offers the transfer.
Disclosure: The agent identifies itself as automated at the top of the call. It is a scored line on the rubric, so a regression in the prompt shows up as a failing check rather than a complaint.
Tool calls: Read and write against CRM, order management, ticketing and scheduling systems during the conversation. Writes are allow-listed per call type and every one is recorded in the call's audit trail.
Failure: If recognition, the model or a downstream system stalls past its budget, the call falls back to a human or to a defined message. It does not sit in silence, which is the worst outcome available on a phone call.
Recording: Consent captured according to the rules of the jurisdiction the number sits in. Retention is set by you, per call type, and deletion is honoured through to the transcripts and scores derived from the audio.

Language

Hindi, English, and whichever language the caller switches into halfway through.

Most speech stacks are tuned on American English telephony audio and then asked to pick one language per call. Neither assumption survives an Indian support queue.

Code-switching: Handled inside the sentence, not by choosing a language for the call. Hindi grammar carrying English nouns and an English number is the ordinary case on an Indian line, not an edge case, and a per-call language choice loses half of it.
Numbers and identifiers: An order number, an amount or an OTP read aloud inside a mixed sentence is where recognition most often fails — and it is usually the thing the call turns on. Digits are read back to the caller for confirmation rather than assumed correct.
Names and places: Indian personal names and locality names are the hardest tokens for a general model, and mishearing one costs the call. They are resolved against your own customer and address records instead of guessed from audio alone.
Accent and region: The same language is a different acoustic problem by region, and a national queue takes several within an hour. Coverage is measured on your recordings before the agent takes a live call, not inferred from the language name.
Following the caller: The caller sets the language and may change it mid-call. The agent follows, rather than holding whichever language was picked at the start.
What is in scope: Hindi, English, Tamil, Telugu, Marathi, Bengali, Kannada, Malayalam, Gujarati and Punjabi. Which of them clear the bar for a given queue is answered by running that queue's own audio — a published list is not evidence.

Integration

It joins your telephony. It does not replace it.

Numbers, routing rules, queues and reporting stay where they are. Rubricon takes a leg of the call.

Layer	How Rubricon connects
Telephony	SIP trunk or a CPaaS media stream. The agent is a destination in your existing routing, so failover and reporting keep working as they do today.
Contact centre	Transfer to a queue or a named agent, with call variables and the transcript passed across on the transfer.
Systems of record	REST or database reads for lookups; allow-listed writes for status changes, bookings and ticket creation.
Knowledge	Your policy documents, tariff sheets and scripts, indexed for retrieval and re-indexed when they change.
Output	Recording, diarised transcript, structured call outcome, tool-call log and the rubric score — pushed to your warehouse or pulled from an API.