AuraVoice
Sovereign Cognitive Telephony Platform
On-premises, GPU-accelerated voice automation for citizen help-desks and enterprise contact centres, with speech in 11 Indian languages plus English, a visual Agent Studio and SIP, PBX and cloud-telephony integration.
- Fully local speech pipeline: no audio or transcript leaves the server
- 11 Indian languages plus English, with Hinglish and regional dialect handling
- Agent Studio: drag-and-drop call flows with live voice testing
- SIP trunks, PBX, Exotel, Twilio, Plivo and WebRTC callers
Overview
Rigid IVR menus and English-first voice bots fail callers who speak Hindi, regional languages or a mix of both, and cloud voice services send citizens’ audio outside the organisation. AuraVoice™ is an on-premises, GPU-accelerated voice-automation platform built for high-volume citizen help-desks and enterprise contact centres: speech recognition, the call language model and speech synthesis all run on the customer’s own GPU, so no audio or transcript leaves the server.
Agents are designed on a drag-and-drop canvas, answer from each bot’s own documents, and connect to SIP trunks, existing PBXs, cloud telephony providers and web callers. Every call is recorded with a PII-masked transcript, outcome and latencies, feeding analytics and SLA reports. Each tenant can be fully white-labelled.
Key capabilities
-
Agent Studio
Drag-and-drop call-flow designer with 14 block types, per-agent language, voice, model and knowledge, live voice or text testing, versioned publishing and rollback.
-
Real-time voice calls
One call session for every channel with barge-in, keypad digits, silence timeouts, transfers and hang-up.
-
Indian-language speech
IndicConformer speech recognition and IndicF5 speech synthesis on the local GPU; Whisper language detection on the first utterance.
-
Hinglish and regional dialects
Dialect decoding for Bundeli, Bagheli, Malwi, Nimadi and Hinglish.
-
Hybrid dialogue
Scripted flows (grievance registration and status, e-Seva) first; anything else goes to the local LLM, grounded in the bot's documents and answered in the caller's language.
-
Grounded knowledge
Per-bot document knowledge base (PDF / text); answers are phrased from matching passages.
-
Responsible-AI guardrails
PII masking (Aadhaar, phone, e-mail), prompt-injection firewall, AI risk register, sentiment analysis and confidence-gated escalation.
-
Voice biometrics
Caller voice authentication with an anti-spoof scoring hook.
-
Transfer with summary
On transfer the officer receives the collected details, the caller's last words and sentiment.
-
Call records and analytics
Stored call records with PII-masked transcripts and summaries; containment, outcomes, languages, intents, latency percentiles, rising call reasons and SLA reports.
-
Outbound campaigns
Number lists called through Asterisk, Twilio or Plivo with concurrency limits, retries and calling hours.
-
White-labelling
Client name, logo, theme, colours, typography and bot persona per tenant.
Architecture
AuraVoice runs as Docker services on a GPU host: an nginx gateway with TLS and WAF rules, a Go orchestrator for calls, flows and APIs, a local speech service (IndicConformer ASR, IndicF5 TTS), Ollama for the call LLM, Asterisk for SIP, PBX and WebRTC, and PostgreSQL and Redis. Speech and language models share the host GPU, so no audio or transcript leaves the server.
-
Callers and channels
- SIP trunk / PBX / PRI gateway
- Exotel, Twilio, Plivo
- WebRTC web callers
-
Gateway
- nginx :443 TLS 1.2 / 1.3
- URI-level WAF rules
- 50 MB upload cap
-
Telephony
- Asterisk (SIP 5060/5061, RTP)
- AudioSocket to the core
- AMI control
-
Orchestrator
- Go core: REST + WebSocket
- Agent flows and call sessions
- Guardrails, RAG, call records
-
Speech and language (GPU)
- IndicConformer ASR
- IndicF5 TTS · Kokoro (English)
- Whisper language ID
- Ollama call LLM
-
Data
- PostgreSQL
- Redis (512 MB LRU)
Technical specifications
| Recognition | IndicConformer (22 languages recognised), Whisper for language ID and English |
|---|---|
| Synthesis | IndicF5 (11 Indian languages); Kokoro-82M Indian-English voices; Indic Parler-TTS optional |
| Full voice (ASR + TTS) | Hindi, Marathi, Bengali, Assamese, Gujarati, Punjabi, Odia, Tamil, Telugu, Kannada, Malayalam, English |
|---|---|
| Dialects | Bundeli, Bagheli, Malwi, Nimadi, Hinglish |
| Default call LLM | gemma3:12b on Ollama (local GPU); optional hosted LLM; rule-based fallback |
|---|
| Reference GPU memory | ~15–18 GB on one NVIDIA L4 (24 GB) for the full pipeline |
|---|
| Providers | Asterisk (SIP trunks, PBX, PRI/E1, WebRTC), Exotel, Twilio, Plivo, any WebSocket media gateway |
|---|---|
| Audio | 8 kHz PCM / μ-law and 16 kHz PCM |
| Runtime | Docker Compose; Kubernetes manifests for production |
|---|
| Stores | PostgreSQL, Redis |
|---|
| Access | JWT sign-in with lockout after 5 failures, bcrypt, Admin / Editor / Viewer roles, persistent audit log |
|---|---|
| Gateway | TLS 1.2 / 1.3, WAF rules (XSS, SQL injection, path traversal), per-IP rate limiting |
Deployment
-
On-premises GPU host
Single GPU host with Docker Compose; all speech and language models on the local GPU.
-
Kubernetes (production)
Multi-node production deployment with encrypted storage and an air-gapped inference path.
Requirements
| Staging / demo | Ubuntu 22.04 / 24.04 or Debian 12; 4 vCPU / 16 GB RAM; 1× NVIDIA L4 (24 GB) or better; 150 GB SSD |
|---|---|
| Production | Ubuntu 24.04 LTS (CIS-hardened); 16+ vCPU / 64+ GB per node; 2× L40S / A100-class GPUs per node; encrypted SSD volumes |
| Network | 80 and 443 inbound; SIP 5060/5061 and RTP at the media gateway; outbound HTTPS during install |
| Capacity | Scales out with additional GPUs; sizing is agreed per deployment |
Supported integrations
-
Asterisk, Cisco, Avaya and other PBXs
SIP / PBX
Through the bundled Asterisk over SIP trunks, PBX or PRI/E1 gateways.
-
Exotel, Twilio, Plivo
Cloud telephony
Webhooks and media streams with token and signature checks.
-
WebRTC
Web callers
Browser softphones via Asterisk.
-
CRM / e-Seva
Citizen services
Grievance ticket and e-Seva application adapters.
Security & governance
-
Local speech and AI
Recognition, reasoning and synthesis run on the host GPU; no audio or transcript leaves the server.
-
Prompt-injection firewall
Blocks prompt injection and jailbreak attempts.
-
PII masking
Aadhaar, phone and e-mail masked; the LLM only sees masked text.
-
Grounded answers
Answers come from the bot's documents, with confidence-gated escalation.
-
Access control and audit
JWT, bcrypt, role-based access and a persistent audit log.
-
Carrier security
Secret webhook tokens and Twilio signature verification.
Use cases
-
Public citizen services
State CM helplines and grievances
Automates complaint registration, revenue record enquiries and farmer benefit status in rural dialects without queue delays.
-
BFSI and fintech
Multilingual collections and fraud verification
Conversational loan reminders and real-time transaction verification without caller biometric leakage.
-
Disaster and emergency services
Emergency triage and alert dispatch
Triage of distress calls on on-premises clusters during weather events and outages, without cloud dependency.
Documentation & developer resources
Telephony Onboarding
Language Support
Architecture
Open the AuraVoice documentation
Downloads
- AuraVoice — Technical & Customer Deck
AuraVoice technical and customer briefing (PowerPoint, 10 slides).
Frequently asked questions
Which languages does AuraVoice support?
Full voice in 11 Indian languages — Hindi, Marathi, Bengali, Assamese, Gujarati, Punjabi, Odia, Tamil, Telugu, Kannada and Malayalam — plus English, with dialect handling for Bundeli, Bagheli, Malwi, Nimadi and Hinglish.
Does any audio leave our infrastructure?
No. Speech recognition, the language model and speech synthesis run on the local GPU.
Which telephony can it connect to?
SIP trunks, existing PBXs and PRI/E1 gateways through the bundled Asterisk, WebRTC callers, and Exotel, Twilio or Plivo.