AI voice agents, IVR, and telephony gateways engineered for 500K calls a day. We cut one partner's per-minute cost by 60% and moved their AI applications across without changing a line of code.
Voice is the channel your customers reach for when something matters and the one most enterprises still rent by the minute. Here's how we think about the decisions that shape a voice build, before anyone writes a call flow.
What is programmable voice?
Programmable voice is telephony your software controls. Rather than configuring a fixed call flow in a vendor dashboard, your application decides what happens on every call: which prompt plays, what the caller said, where the call routes, when an AI agent handles it, and when a person takes over.
Underneath sit four layers: SIP trunks that connect you to the phone network, a media layer that moves audio in real time, speech recognition and synthesis that turn that audio into language and back, and the business logic that ties all three to your systems. Own those layers and voice becomes a product surface you can change weekly instead of a contract you renew annually.
Managed platform or self-hosted infrastructure?
Managed voice platforms are the right call at low volume. They handle carrier relationships, compliance, and scaling, and you pay a per-minute premium for that. The premium stops making sense somewhere north of a few hundred thousand minutes a month, where the same traffic on a direct carrier trunk costs a fraction of the platform rate.
The barrier to moving is rarely price. It's that applications get written against a specific provider's protocol, so leaving means rewriting every voice bot. We solve that by replicating the protocol your applications already speak, which is exactly how one migration off Exotel shipped with zero changes to the AI application code.
What latency budgets actually demand
A voice conversation feels natural when one-way audio latency stays under roughly 200 milliseconds. Every stage spends from that budget: media transcoding, network hops, speech recognition, the language model, and synthesis. Miss it and callers talk over the agent, which shows up in your completion rates long before it shows up in a dashboard.
Holding the budget is engineering discipline rather than a vendor choice. Custom codecs instead of general-purpose media tooling, streaming recognition instead of batch, pre-generated audio for anything you already know you will say, and per-stage latency tracking so you know which hop to fix when the number drifts.
Multilingual voice at national scale
For programs serving Indian citizens, language coverage decides participation. Commercial speech APIs handle a handful of Indian languages well and price the rest out of reach at scale, which is how survey programs end up English-only and reaching a minority of the people they exist to serve.
Bhashini changes that arithmetic. Building on its Dhruva inference APIs gave one welfare program recognition, translation, and synthesis across 14 Indian languages in a single pipeline under two seconds end to end, with no human translators in the loop. The pipeline stays provider-agnostic, so adding a market is a routing change rather than a rebuild.
What programmable voice changes for the business
Lower cost per minute
Buying minutes directly from a carrier instead of through a platform markup is where the saving comes from. On one gateway we built, that moved per-minute cost from Rs 0.50 to Rs 0.20.
Capacity that follows demand
Voice automation absorbs peak volume without a hiring cycle. A campaign that would take a floor of agents a week runs overnight across the same infrastructure.
Consistency on every call
The same questions, the same disclosures, and the same routing rules on call 500,000 as on call one, with a transcript for every conversation.
Infrastructure you own
Self-hosted gateways keep audio, recordings, and call data inside your own estate, in the region you choose, under your retention policy.
Our programmable voice capabilities
Voice AI agents
Conversational agents that answer, ask, and act on live calls. Built as state machines with explicit phase transitions, so the agent follows your script instead of improvising its way off it.
IVR systems
Menu-driven and natural-language IVR, from a five-option keypad tree to intent routing that skips the tree entirely. Both run against the same call flow engine and the same analytics.
Inbound and outbound calling
Queue-backed campaign dialing for outbound, capacity-aware routing for inbound, and retry policies that respect calling windows, consent, and do-not-disturb registries.
SIP and telephony integration
Direct carrier SIP trunks, PJSIP and Asterisk PBX configuration, and a pluggable trunk provider interface where switching between a carrier, Exotel, or Twilio is a config change.
Real-time speech-to-text
Streaming transcription on the live call rather than after it, so intent, sentiment, and compliance triggers land while the caller is still on the line.
Text-to-speech
Neural voice synthesis with variable substitution for names, amounts, and dates. Static prompts are pre-generated through queues so known content plays with no synthesis wait.
Multilingual voice experiences
Speech recognition, machine translation, and synthesis across 14 Indian languages through Bhashini, with commercial providers for other markets behind the same pipeline.
Call recording and analytics
Recordings, transcripts, and per-stage latency metrics stored in your own infrastructure, with dashboards covering completion rates, drop-off points, and agent performance.
Workflow automation
Calls that read and write your CRM, ticketing, billing, and ERP systems through queue-backed jobs, so a slow downstream system delays a record rather than dropping a call.
Human handoff
Defined escalation triggers on confidence, sentiment, and regulated topics, transferring the live call to a human with transcript, intent, and caller context already attached.
Solutions we build
Customer support automation
Voice agents that resolve balance checks, order status, and account changes end to end, and escalate the rest to a human with the context already gathered.
Survey and feedback calls
Automated outbound surveys in the respondent's own language, writing structured responses straight into your data store without a human transcription step.
Interview and recruitment automation
Structured screening interviews conducted by an AI interviewer, with evidence-anchored scoring so hiring teams get comparable candidate data instead of recordings to rewatch.
Appointment reminders
Confirmation, reschedule, and cancellation calls driven by your booking system, with the outcome written back automatically and no-shows flagged for follow-up.
Collections and follow-ups
Payment reminders and promise-to-pay capture with calling-window and consent rules enforced in the flow, and a transcript trail for every regulated disclosure.
Public service and government outreach
Citizen-facing voice programs at state and national scale, built on sovereign speech services and multilingual pipelines so language is never the barrier to participation.
Contact center modernization
Replacing legacy IVR and per-minute platform contracts with self-hosted infrastructure, then layering AI agents, real-time transcription, and analytics on top of it.
Our voice technology stack
SIP
PJSIP and Asterisk PBX against carrier trunks including BSNL, with registration, channel management, and failover engineered for sustained concurrency.
WebRTC
Browser and app calling over LiveKit's SFU architecture, used in production for live AI video interviews at roughly 150ms round-trip latency.
Voice AI models
LangGraph for multi-agent orchestration and GPT-4o class models for language, wrapped in state machines that enforce phase discipline a flat prompt loop cannot.
Bhashini
India's sovereign language stack through the Dhruva inference APIs: automatic speech recognition, neural machine translation, and synthesis across 14 Indian languages.
STT and TTS providers
Provider-agnostic speech routing with per-stage latency tracking and error isolation, so a degraded upstream service fails one stage instead of the whole call.
Telephony gateways
Custom media gateways with real-time transcoding between PSTN mu-law RTP and PCM audio, processing each 20ms chunk in under a millisecond.
Self-hosted voice infrastructure
NestJS services, PostgreSQL, Redis, and BullMQ on Docker across cloud or on-premises, with distributed deployments where VPN and network constraints demand them.
Real-time streaming architecture
WebSocket audio streaming, RTP/UDP media paths, and Prometheus instrumentation on every hop, because a latency budget you cannot measure is one you cannot hold.
Why enterprises choose us for voice
We have replaced a legacy provider in production
We reverse-engineered Exotel's WebSocket protocol and built a drop-in gateway that existing AI voice applications connected to without a single code change. Migration became a config change, not a rewrite.
Cost optimization and infrastructure ownership
A bill of over Rs 10 million per month at Rs 0.50 per minute became Rs 0.20 per minute on infrastructure running at roughly $400 per month. The trunk provider interface keeps that switchable.
Multilingual AI expertise
A unified speech recognition, translation, and synthesis pipeline across 14 Indian languages on Bhashini, running under two seconds end to end with no human translators in the loop.
High-volume call processing
Systems engineered for 500,000 calls per day, roughly 3,000 to 6,000 concurrent channels, with least-connections routing across node pairs and session state that survives restarts.
Enterprise-grade security and reliability
One-way audio latency under 200ms, packet loss under 1%, no single points of failure, and recordings and call data held inside infrastructure you control.
Featured case studies
Three production voice systems we designed, built, and deployed. Each links to the full engineering write-up, including the constraints and the decisions behind the numbers.
Telephony gateway
Replacing Exotel with a self-hosted voice AI gateway
A social-impact foundation running AI voice surveys across India was locked into Exotel at Rs 0.50 per minute. We built an Asterisk-based gateway that speaks Exotel's exact WebSocket protocol and connects straight to BSNL SIP trunks, then split the gateway and PBX across servers to work around a VPN that blocked outbound internet access.
A multilingual voice AI pipeline on Bhashini for public welfare surveys
A public welfare program needed citizen surveys across multiple Indian states, but its English-only system excluded most of the target population. We built a single speech-to-text, translation, and text-to-speech pipeline on Bhashini's Dhruva APIs, pre-generating static survey audio through queues and synthesizing only the personalized parts live.
High-volume hiring needed structured interviews without structured interviewer time. We built a multi-agent platform where an AI avatar conducts the interview over live WebRTC video, transcribes it in real time, and scores five behavioral dimensions with every rating anchored to a verbatim transcript citation.
We map call volumes, peak concurrency, current per-minute spend, carrier contracts, and the systems the calls have to talk to. You get a straight answer on whether self-hosting pays back at your volume.
2
Architecture design
Trunk strategy, media path, speech providers, and AI orchestration, sized against your latency budget and your compliance constraints before anyone writes call flow code.
3
Development
Call flows, agent logic, integrations, and the gateway layer built in tight increments, tested against real trunks and real audio rather than a simulator.
4
Deployment
Staged cutover with the old provider still live behind a pluggable trunk interface, so traffic moves in controlled percentages and rolls back in seconds if a metric moves the wrong way.
5
Monitoring and optimization
Prometheus instrumentation on every hop, plus tuning on prompt design, routing rules, and speech model selection as real call data shows where conversations break down.
Frequently asked questions
What is programmable voice?
Programmable voice is telephony you control with code. Instead of a fixed call flow configured in a vendor dashboard, your application decides what happens on every call: which prompt plays, what the caller said, where the call routes, when an AI agent answers, and when a human takes over. The building blocks are SIP trunks for carrier connectivity, a media layer that moves audio in real time, speech-to-text and text-to-speech, and the business logic that ties them together.
Can you replace our existing voice provider without rewriting our applications?
Yes, and we have. For a social-impact foundation we built a self-hosted gateway that speaks Exotel's exact WebSocket protocol, so their existing AI voice applications connected to our gateway with zero code changes. The migration was a configuration change rather than a rewrite. The same approach applies to most providers: we replicate the protocol your applications already speak, then swap the carrier underneath.
How much can we save by self-hosting voice infrastructure?
It depends on your call volume and your current per-minute rate, because the saving comes from buying minutes directly from a carrier instead of through a platform markup. On the gateway we built, per-minute cost dropped from Rs 0.50 to Rs 0.20, a 60% reduction against a bill of over Rs 10 million per month, with the replacement infrastructure running at roughly $400 per month. Volume is what makes the economics work: below a few hundred thousand minutes a month, a managed provider is often the rational choice, and we will tell you that on the first call.
How many concurrent calls can the systems you build handle?
The gateway we deployed is engineered for 500,000 calls per day, roughly 3,000 to 6,000 concurrent channels, with horizontal scaling across gateway and PBX node pairs and least-connections routing between them. Capacity is a function of media servers, trunk channels, and the AI services behind the call, and we size each layer against your peak concurrency rather than your daily total.
Which languages do your voice systems support?
For Indian languages we build on Bhashini's Dhruva inference APIs, which gave one public welfare program speech recognition, neural machine translation, and speech synthesis across 14 Indian languages in a single pipeline running under two seconds end to end. For other markets we integrate commercial speech providers. The pipeline is provider-agnostic by design, so language coverage is a routing decision rather than an architectural one.
What latency should we expect from an AI voice agent?
Conversation feels natural when one-way audio latency stays under roughly 200 milliseconds and the agent's first response lands within a second or two. Our telephony gateway holds one-way audio latency under 200ms with packet loss under 1%, and the interview platform we built runs audio and video round-trip at about 150ms over WebRTC. Getting there is mostly budget discipline: custom transcoding rather than generic media tooling, streaming speech recognition rather than batch, and pre-generated audio for anything you already know you will say.
Do we need to move off our current telephony provider to work with you?
No. We build against a pluggable trunk provider interface, so a carrier SIP trunk, Exotel, and Twilio are interchangeable behind the same application code and switching between them is a configuration change. Many partners start by keeping their existing provider and adding AI agents, routing, or analytics on top, then move traffic to a direct trunk once the economics justify it.
How do you handle handoff from an AI agent to a human?
Handoff is part of the call flow design, not an afterthought. We define the triggers up front, including explicit caller requests, low confidence in the agent's understanding, sentiment shifts, and regulated topics the agent should not handle, then transfer the live call to an agent queue with the transcript, intent, and caller context attached so the person picking up does not start from nothing.
Can voice systems integrate with our CRM and internal systems?
Yes. That integration is usually where the business value sits. Calls trigger and read from your CRM, ticketing, ERP, or billing systems through their APIs, so an outbound reminder knows the appointment, a collections call knows the balance, and a completed survey writes straight into your data store. We build these as queue-backed workflows so a slow downstream system delays a record rather than dropping a call.
How do you handle call recording, compliance, and data residency?
Recordings and transcripts are stored in infrastructure you own, in the region you choose, with retention and access controls set by your policy. Self-hosting is what makes that possible: audio never transits a third-party platform unless you decide it should, and speech processing can run on sovereign services such as Bhashini or on models deployed inside your own network for regulated workloads.
Bring us your call volumes and your current per-minute spend. We'll tell you what we'd build, what it would cost to run, and whether self-hosting pays back at your scale.
