Voice AI Implementation: From SIP Trunking to NLU Pipeline

SIP Trunking for Voice AI in India – Summary

SIP trunking connects your voice AI to the Indian PSTN, but India’s regulatory and network environment differs from Western markets.

SIP Provider Options

• Exotel
• India-first cloud telephony platform
• Strong Indian DID/local number support
• Built-in IVR and compliance with TRAI/Indian telecom rules
• Ideal when you need Indian numbers and regulatory compliance out of the box
• Ozonetel
• Contact center–focused platform with integrated SIP trunking
• Call recording, analytics, and CRM integrations
• Common choice for Indian enterprises and support centers
• Twilio
• Global CPaaS with expanding Indian number coverage
• Good when you also need international connectivity
• Fits teams already using Twilio for SMS/WhatsApp or global voice

Key Technical Considerations

• Call routing & failover
• Configure primary and secondary SIP trunks
• Aim for 99.9%+ uptime with geo-redundant trunks (e.g., Mumbai + Bangalore DCs)
• Helps mitigate regional Indian telecom outages
• Audio codecs
• Use G.711 (PCMU/PCMA) for maximum compatibility with Indian carriers
• Opus can be used internally for better quality and lower bandwidth, but PSTN calls usually require transcoding to G.711
• DTMF handling
• Many Indian callers still rely on touch-tone IVR menus
• Support RFC 2833 (out-of-band) DTMF and in-band detection for older/legacy networks
• Concurrent call capacity
• Size trunks for peak concurrent calls (typical mid-size: 50–100 channels)
• Incoming call costs: roughly ₹0.50–₹1.50 per minute
• Telephony minutes often become the largest recurring cost component
• Number porting & DID
• Coordinate porting from existing IVR/PRI setups with your SIP provider
• Indian number porting typically takes 5–7 business days
• Plan cutover and testing windows carefully to avoid downtime

Integration Architecture

A typical India-focused voice AI deployment uses:

1. SIP trunk (Exotel/Ozonetel/Twilio) connecting to a media server (Asterisk, FreeSWITCH, or cloud-native SBC/media service).
2. The media server handles:

• Call setup/teardown via SIP (INVITE, BYE, re-INVITE, etc.)
• RTP audio capture and forwarding to the ASR engine
• TTS playback back to the caller over RTP
• Call transfer to human agents (SIP REFER/transfer or via contact center APIs)
• Call recording for QA, analytics, and compliance.

This architecture ensures regulatory compliance, reliable connectivity to the Indian PSTN, and a robust foundation for production-grade voice AI experiences.

The NLU pipeline converts ASR text into structured meaning using two core tasks: intent classification and entity extraction.

Summary of LLM–NLU Orchestration

Routing Logic

• If NLU confidence > 0.85 → handle directly via dialog manager (no LLM).
• If NLU confidence 0.5–0.85 → call LLM to disambiguate or rephrase the user’s question back for confirmation.
• If NLU confidence < 0.5 or the query is open-ended → route to LLM for full response generation.

In production, about 70–80% of calls are fully handled by NLU; 20–30% require LLM, balancing cost and quality.

LLM Integration Patterns

1. RAG (Retrieval-Augmented Generation)

• Use for knowledge-heavy tasks: product info, policies, troubleshooting.
• Flow: retrieve from vector DB → pass retrieved context + query to LLM → generate grounded answer.

1. Structured Output

• LLM returns JSON instead of free text.
• Dialog manager consumes this JSON to trigger actions, e.g.:
• Available appointment slots
• Payment options
• Policy eligibility decisions

1. Guardrails

• Post-process LLM output through a validation layer that:
• Checks for hallucinations / unsupported claims
• Enforces business rules and constraints
• Filters or blocks unsafe / inappropriate content
• Only validated content is sent to TTS.

Cost Optimization Strategies

• Response caching
• Cache LLM answers for common FAQs and reuse when queries are semantically similar.
• Model tiering
• Use smaller/cheaper models (e.g., GPT-3.5, Claude Haiku) for:
• Simple classification
• Rephrasing
• Light disambiguation
• Reserve larger/expensive models (e.g., GPT-4, Claude Opus) for:
• Complex reasoning
• Multi-step workflows
• High-stakes decisions
• Prompt optimization
• Keep prompts short and focused on the task.
• Reuse shared system prompts and templates.
• Strip irrelevant context to reduce tokens, typically saving 30–50% in token costs.

This setup yields a hybrid system where NLU handles the majority of traffic efficiently, while the LLM provides high-quality, controlled responses for ambiguous or complex queries.

Production Deployment Summary for Voice AI (India)

1. Infrastructure Requirements

For Indian deployments handling 500–5000 concurrent calls:

• Compute
• GPU instances (e.g., NVIDIA T4 or A10G)
• Primary use: ASR (Automatic Speech Recognition) and NLU (Natural Language Understanding) inference
• Provision GPU pools sized for peak concurrency; use autoscaling to handle spikes
• CPU instances
• Use for dialog management, business logic, API orchestration, and non-ML services
• Cloud Region & Redundancy
• Primary region: Mumbai
• AWS: ap-south-1
• GCP: asia-south1
• Rationale: lowest latency to Indian telecom networks and major user bases
• Failover region: Bangalore
• Use as DR / active-passive or active-active failover depending on SLA requirements
• Auto-scaling Strategy
• Use horizontal pod autoscaling (HPA) based on:
• Concurrent call count
• ASR/LLM queue depth
• GPU/CPU utilization thresholds
• Align capacity planning with predictable peak windows:
• 10:00 AM – 1:00 PM IST
• 3:00 PM – 7:00 PM IST
• Redis Usage
• Session state management
• Store per-call context, dialog state, and short-lived metadata
• Response caching
• Cache frequent NLU/LLM responses and static TTS segments where applicable
• Real-time analytics aggregation
• Increment counters and metrics in Redis for low-latency dashboards and alerting

2. Monitoring Dashboard Requirements

A production-grade voice AI system needs real-time observability across four dimensions:

• Call Metrics
• Concurrent active calls
• Average and percentile call duration (P50/P90/P99)
• Queue depth (waiting calls / pending ASR-NLU jobs)
• Abandonment rate (calls dropped before resolution or agent transfer)
• Quality Metrics
• ASR Word Error Rate (WER)
• NLU intent accuracy and confusion matrix
• Containment rate: % of calls resolved without human agent transfer
• Latency Metrics (per turn and end-to-end)
• ASR processing time
• NLU inference time
• LLM response generation time
• TTS generation time
• Total round-trip latency (user speech start → synthesized response start)
• Business Metrics
• First-Call Resolution (FCR) rate
• Customer satisfaction (CSAT) from post-call surveys
• Cost per conversation / per resolved issue

Dashboards should support:

• Real-time views (sub-minute refresh)
• Historical trends (daily/weekly/monthly)
• Threshold-based alerts (PagerDuty/Slack/Email) on key SLOs

3. Continuous Improvement Loop

Voice AI performance improves via a structured feedback cycle:

1. Call Recordings Review

• Sample 5–10% of calls weekly
• Identify:
• ASR misrecognitions (accent, noise, code-mixing issues)
• NLU intent failures and misclassifications
• Dialog design gaps (confusing flows, dead-ends)

1. NLU Retraining

• Monthly retraining using production call data
• Prioritize:
• Low-confidence predictions
• Misclassified examples
• New phrases and colloquialisms from real users

1. Prompt Tuning (LLM)

• Refine system and tool prompts based on:
• Escalation reasons
• CSAT feedback
• Containment vs. transfer patterns

Explore more from our AI solutions library:

• Connecting Claude & GPT-4 to Enterprise Tools via MCP — Integrate your voice AI pipeline with CRM, ERP, and internal tools using Model Context Protocol.
• Document Ingestion Pipeline for Enterprise Knowledge Bases — Feed your voice AI with company knowledge by building a production document ingestion pipeline.