Back to blog
AI Infrastructure

Twilio + Exotel AI Voice Stack: Building a Production-Grade Dual Telephony Architecture

Learn how to build a production-grade AI voice system using Twilio (UK/EU) and Exotel (India) with GPT Realtime, Gemini Live, Deepgram, and modern TTS engines.

Anshul Mittal29 March 2026

Building a Production-Grade Dual-Telephony AI Voice Stack with Twilio + Exotel

Published by Frontly Voice AI · Technical Deep Dive · 10 min read

Voice AI systems have matured beyond demos. Today, building a production-grade voice stack requires careful choices across telephony, AI processing, integrations, latency, and compliance — especially when deploying across regions like India and the UK.

At Frontly, we use a dual-telephony architecture combining Exotel for India and Twilio for the UK/EU with a shared AI backend. This gives us stronger reliability, regional compliance, and lower latency than a single-provider setup.


Why Dual Telephony?

Most AI voice systems rely on one telephony provider. That works — until it doesn’t.

Latency spikes, downtime, carrier issues, or regulatory constraints can quickly break a production experience. A voice system is only as strong as the network layer beneath it.

Our principle: Use the best telephony provider for each market, while keeping the AI backend shared and modular.
  • Exotel for India: strong TRAI alignment, local DID numbers, and low latency from AWS Mumbai
  • Twilio for UK/EU: mature media streaming, Ofcom-friendly number support, and strong European routing

The 3-Tier Architecture

Tier 1: Telephony Layer

Both Exotel and Twilio connect to the AI backend through WebSocket streaming, not REST polling.

This matters because voice conversations are real-time systems. Audio must move continuously in both directions without repeated request-response overhead.

  • Persistent bidirectional audio stream
  • Lower latency than HTTP polling
  • More natural turn-taking in conversation
Technical note: REST can add avoidable latency for every chunk of audio. WebSocket is what makes real-time voice feel like a real phone call instead of a stitched-together workflow.

Tier 2: AI Processing Layer

Incoming telephony audio typically arrives as an 8kHz μ-law stream. Modern STT and TTS systems often require higher sample rates, so the pipeline must handle format conversion carefully.

1. Resampling

Audio is resampled from 8kHz to 16kHz or 24kHz so it can be processed by modern speech engines.

2. Speech-to-Text (STT)

  • Deepgram Nova 2/3 for fast, production-grade transcription
  • ElevenLabs Scribe as an alternate high-accuracy option

3. Language Model (LLM)

  • GPT Realtime
  • Gemini Live

This layer interprets the transcript, applies system prompts, uses contextual business data, and determines the next response.

4. Text-to-Speech (TTS)

  • ElevenLabs
  • Sonic 3
  • Sarvam AI (India workflows)
  • Chirp 3 (UK English workflows)
Important: Chirp 3 may output at 24kHz, which means the response audio must be downsampled back to 8kHz μ-law before telephony delivery. This conversion step is operationally important and often overlooked.

Tier 3: Integration Layer

This is where conversations become real business actions.

  • CRM writes to Salesforce, LeadSquared, Zoho, or HubSpot
  • Calendar booking through Google Calendar or vertical-specific systems
  • WhatsApp confirmations via Meta APIs
  • Webhook callbacks for downstream operational workflows

Without this layer, a voice system only talks. With it, the system actually completes work.


Infrastructure Design: India vs UK

Component India Node (ap-south-1) UK Node (eu-west-2)
Telephony Exotel WebSocket Stream Twilio Media Streams
STT Deepgram Nova 2/3 (Mumbai) Deepgram Nova 2/3 / ElevenLabs Scribe (EU)
LLM GPT Realtime via Azure OpenAI (India) GPT Realtime / Gemini Live via Azure (UK South)
TTS ElevenLabs / Sonic 3 / Sarvam AI Chirp 3 / ElevenLabs
Storage AWS S3 ap-south-1 (encrypted) AWS S3 eu-west-2 (encrypted)
Routing AWS Route 53 geo-routing Cloudflare geo-routing / failover

Latency Breakdown

STT
~280ms
LLM
~420ms
TTS
~180ms
Total Pipeline
~880ms

With streaming TTS, the user perceives the first response chunk earlier, which can bring effective conversational latency closer to ~600ms.


Frequently Asked Questions

Why use WebSocket instead of REST for telephony integration?

Because real-time voice is a streaming problem. REST adds repeated HTTP overhead and unnecessary delay. WebSocket maintains a persistent connection and allows audio to move continuously in both directions.

How do you handle network failover between India and UK nodes?

Route 53 health checks can monitor service availability and latency thresholds, while regional routing and Cloudflare-based failover can redirect traffic when a node degrades.

What is the end-to-end latency of the full voice AI pipeline?

Under normal load, roughly ~880ms total, with perceived latency reduced through streaming response techniques.


Key Takeaways

  • Dual telephony improves resilience, compliance, and regional performance
  • WebSocket streaming is essential for low-latency voice systems
  • Telephony audio format conversion is a real engineering concern, not a minor detail
  • Infrastructure choices in India and the UK should be region-specific
  • The winning voice stacks are modular and operationally optimized, not just model-driven

Final Thoughts

The biggest shift in Voice AI is that the problem is no longer “Can the AI talk?”

The real problem now is: Can the full system perform reliably in production across geographies, telecom layers, and business workflows?

The teams that win in this category will not just choose good models. They will design better infrastructure.

Want a Technical Deep Dive?

Talk to the Frontly engineering team to understand how to deploy a production-grade voice AI system for your use case.