AuraVoice v7.4: Sovereign, model-agnostic Voice AI Agent platform for Indian public services and enterprises

Citizens and customers in India expect to be served in their own language, at any hour, on the channel they already use: a phone call, a website or WhatsApp. Organisations that serve them need more than a chatbot. They need an AI agent platform they can govern: one that keeps data in India, works with the language models they choose, measures its own quality, and always lets a person take over.

AuraVoice is VedaCLOUD.AI’s omnichannel Voice AI Agent platform. Inbound voice agents, outbound calling campaigns, a website assistant, WhatsApp conversations and live assistance for human agents run on one orchestration layer, with one set of agents, one knowledge base, one audit trail and one set of rules. AuraVoice is our own product: designed, built and versioned by VedaCLOUD.AI, and deployable on premises, in an Indian cloud region, or both.

This article describes what the platform does today, release 7.4.

---

At a glance#

Area	What AuraVoice provides
Channels	Inbound and outbound voice (SIP trunks, PBX and PRI gateways, PSTN carriers and CPaaS), WebRTC browser calling, an accessible website assistant, two-way WhatsApp, SMS and e-mail notifications, live agent assist
AI models	Model-agnostic: several providers each for the language model, speech recognition and voices, switched by configuration, with configured secondary providers and automatic failover
Indian languages	English, Hindi and the major Indian languages, including accented and code-mixed (Hinglish) speech; 22 languages through Bhashini; language detection or caller choice; keypad (DTMF) fallback
Knowledge	Documents, spreadsheets and approved web pages, with approval, versioning, review dates and source attribution in every answer
Workflows	No-code agent studio, AI-assisted agent creation, API calls to CRM, ERP, ticketing and grievance systems, with validation, confirmation and audit
Governance	Automated-assistant disclosure, human fallback after two failed turns, safety triggers, callbacks, approval workflows, evaluation before publishing
Security	Role-based access, single sign-on (OpenID Connect), short-lived tokens, caller verification, credential vault with OAuth 2.0, signed and replay-protected requests, tool-call allowlists
Evidence	Per-turn latency and model provenance, semantic-cache statistics, availability, failover drills, AI-quality evaluations and speech-accuracy tests, exported as JSON and Prometheus metrics and streamed to a SIEM
Data	Processed and stored in India; data-residency view of every component; erasure on request; retention controls

---

1. One platform, every channel#

AuraVoice agents are built once and answer everywhere:

• Phone calls: inbound numbers on SIP trunks, an existing PBX or contact-centre switch (through the bundled Asterisk), PRI through standard media gateways, and Indian and global carriers such as Exotel, Twilio and Plivo; any media gateway can also connect over WebSocket. Callers can interrupt the agent (barge-in); silence, retries and unclear answers are handled by the flow; keypad input is accepted wherever speech is.
• Outbound campaigns: reminders, feedback and outreach calls with concurrency, retries, calling windows and per-number outcomes linked to call records.
• Website assistant: a single script tag puts any agent on a website as a chat assistant designed to WCAG 2.1 Level AA: fully keyboard operable, announced to screen readers, high contrast, resizable text, reduced motion and language-tagged messages.
• WhatsApp: the same agents hold two-way WhatsApp conversations through the WhatsApp Business Cloud API or Twilio, with the same flows, knowledge and rules.
• Notifications: flows send SMS (any Indian SMS gateway with DLT templates), WhatsApp messages and e-mail: ticket numbers, appointment details, links.

Every conversation, on every channel, is recorded as a call record with its transcript (personal identifiers masked), outcome, collected details, path through the flow, summary and timings.

2. Model-agnostic by design#

AuraVoice separates the conversation logic from the models that power it. Each of the three model roles has several interchangeable providers:

• Language models: self-hosted open models on the platform’s own GPU (for example Llama and Gemma through Ollama), OpenAI, Azure OpenAI, Google Gemini, Sarvam, and any OpenAI-compatible service, including Indian and sovereign models such as Krutrim and Sarvam-M, and private vLLM or LiteLLM deployments.
• Speech recognition: self-hosted AI4Bharat IndicConformer, Bhashini, Sarvam (Saarika, code-mixed speech), Google Cloud Speech, Microsoft Azure Speech, OpenAI and ElevenLabs.
• Voices: self-hosted Indian voices (AI4Bharat IndicF5 and Indic FastPitch, Kokoro for Indian English), Bhashini, Sarvam Bulbul, Google Chirp 3 HD and Neural2, Azure Neural Indian voices, ElevenLabs and OpenAI, with per-agent voice, pitch, pace and speaking-style prompts.

Switching the platform from one provider to another is a configuration change, applied from the next call, with no change to any agent and no change to the application. Adopting an Indian-developed or sovereign model requires no re-architecting.

Resilience. For each role an administrator selects a secondary provider. If the primary fails (outage, quota, timeout), the same request is answered by the secondary, and then by the self-hosted engine, so a call never goes silent. Every switch is logged with its reason and switch time, and a failover drill can be run on demand and kept as evidence.

Model governance. Every agent is versioned: drafts are tested and published as numbered versions, calls stay on the version they started with, and any earlier version can be made live again in seconds. Every turn records which provider and model produced the answer, including after a failover, so each interaction has its model provenance.

3. Speed that is measured, not claimed#

Response time on a phone call is the time between the caller finishing a sentence and hearing the agent begin to reply. AuraVoice measures exactly that, for every turn: from the end of the caller’s speech to the first byte of reply audio leaving for the telephony side. Reports show the 50th, 90th and 95th percentiles, separated by how each answer was produced:

• Flow rules: menus, collected details and fixed prompts, with no model involved.
• Semantic cache: AuraVoice recognises a question that has been answered before, even in different words (“How many days of earned leave do I get?” and “How many earned leave days do I get in a year?”), and answers it in milliseconds without calling the language model. The cache covers both the routing decision and the answer, is scoped to each agent version, never matches questions with different numbers or a negation, and is cleared automatically whenever the knowledge base changes.
• Language model: everything else.

The cache hit rate, its latency gain over model answers, and the latency percentiles are shown on the SLA report and exported in machine-readable form, so they can be verified independently on the organisation’s own questions and traffic. Replies are also spoken sentence by sentence and frequently used prompts are cached as audio, so the first words reach the caller as early as possible.

4. Knowledge the agent can stand behind#

Agents answer from the organisation’s own material, not from the model’s memory:

• Sources: PDF, Word, Excel and CSV files, HTML and text, and approved portal or web pages fetched and refreshed on a schedule. Spreadsheets become one entry per row with the column names, so a question finds the row that answers it. Live systems such as databases or CRMs are reached through API calls in the flow.
• Lifecycle: every document version is kept with a SHA-256 integrity hash; new or changed material can require an administrator’s approval before agents use it; review dates retire outdated material automatically. Updating knowledge never requires code changes or redeployment.
• Grounding: answers come only from matching passages. When the documents do not cover a question, the agent says so and the flow continues with its human-authored fallback, rather than improvising. Every answer records the documents it came from, shown in the test console, the call record and the website assistant.

5. Agentic workflows without code#

The Agent Studio is a visual flow designer: messages, questions with validation, keypad menus, intent routing, knowledge answers, free-flowing AI agent steps that collect structured details, conditions, API calls, caller verification, notifications, business hours, language switching, transfers and endings. Agents can be tested by voice or text with the path lighting up live, validated, published and rolled back. Build with AI turns a plain-language description and a few answers into a complete, validated draft agent.

API calls connect to CRM, ERP, helpdesk, ticketing and grievance systems, portals and custom applications over REST with JSON bodies (GraphQL services are called the same way), with values from the conversation and mapping of response fields back into the flow. Actions are controlled:

• Validation of every collected value (numbers, dates, mobile numbers, patterns, choices) before it is used.
• Confirmation: payments, cancellations and other irreversible actions can require the caller’s explicit “yes” before the API is called.
• Caller verification: a verification step (one-time code to the registered number, or a known answer such as a date of birth from the CRM) raises the call’s assurance level; actions and personal data can require a level, because caller ID alone is not authentication.
• Allowlists and least privilege: API calls can only reach allow-listed hosts; credentials live in a vault, are referenced by name, and are sent only to the hosts they belong to; OAuth 2.0 client-credential tokens are obtained and renewed automatically.
• Signed requests: every outgoing call carries an HMAC-SHA256 signature over a timestamp, a nonce and the body, so receiving systems can reject forged and replayed requests.
• Audit: every action, its result and its timing appear in the call’s trace and the audit log.

6. Human agents and supervisors stay in control#

• Context-preserving handover: on transfer, the officer receives the details collected, the caller’s last words and sentiment, and a suggested next action, so the caller is never asked to repeat themselves. Transfers go to phone numbers, extensions or contact-centre queues.
• Agent assist: while a call is live, supervisors and agents see each turn as it happens, with knowledge passages matching the caller’s latest question.
• Callbacks: if no person can take the call, the caller is told a callback has been arranged, and the request appears in a callback queue with the conversation’s context.
• Dashboards and quality: live calls, outcomes, containment, languages, intents, sentiment, busiest hours and rising call reasons; per-call AI analysis with configurable quality scorecards, extracted fields and after-call summaries; call recordings with both sides time-aligned.

7. Built for India’s languages and for accessibility#

AuraVoice understands and speaks English, Hindi and the major Indian languages, and handles accented and code-mixed speech, including Hinglish. Through Bhashini it reaches 22 Indian languages. The caller’s language is detected automatically, limited to the languages the agent serves, or chosen from a keypad menu; messages written in one language are translated for callers in another. Every voice journey offers keypad fallback for callers who cannot or prefer not to speak, and prompts are concise and clear for telephone audio. The website assistant is designed to WCAG 2.1 Level AA.

Speech-recognition accuracy is tested language by language on real recordings: AuraVoice computes word and character error rates per language and recogniser, so the best recogniser can be chosen for each language and the results kept as evidence.

8. Responsible AI, enforced by the platform#

Some rules are too important to leave to each agent’s design. AuraVoice enforces them in the runtime, for every agent:

• Every interaction starts by telling the caller they are speaking with an automated assistant.
• After two failed turns, immediately when the caller asks for a person, or on a safety trigger (a medical emergency, a threat, self-harm), the caller is transferred to a human, or a callback is arranged when nobody is available. Automated loops that deny human fallback are not possible.
• Publishing can require an administrator’s approval and a passing evaluation run of the exact draft being published.
• Voice cloning is off unless specifically authorised; recorded caller audio is never used to train or synthesise voices.
• Organisation data, recordings, transcripts and prompts are not used to train or improve any model.
• Personal identifiers (Aadhaar, phone, e-mail) are masked before they reach language-model prompts, logs, analytics or stored transcripts; a prompt-injection firewall blocks attempts to override the agent’s instructions.

9. Quality you can test before and after every change#

AuraVoice includes an evaluation framework owned by the organisation. Each agent has a versioned evaluation set: real questions with the facts a correct answer must contain, statements it must not make, the documents it must cite, and the conversations that must end with a person. Built-in security probes cover direct and indirect prompt injection, jailbreaks, role-play to bypass rules, attempts to extract other callers’ data, tool-call abuse and harmful requests.

A run plays every case through the real agent, knowledge base and model, exactly as a call would, and scores answer accuracy, made-up answers, correct human escalation and security. Runs are kept per agent version: before acceptance, and after every model, prompt, knowledge or workflow change. Combined with speech-accuracy testing and the latency measurements above, this gives a complete, repeatable quality record.

10. Security, privacy and compliance#

• Access: role-based access (Administrator, Editor, read-only Viewer), single sign-on with the organisation’s OpenID Connect identity provider (authorization code flow with PKCE, signature-verified identity tokens, roles from groups), short-lived console sessions, lock-out after failed sign-ins, and an audit log of every administrative action.
• Transport and storage: TLS 1.2/1.3 at the gateway with web-application-firewall rules and rate limiting; carrier webhooks protected by tokens and signatures; secrets in restricted files and a credential vault; deployable on encrypted storage.
• Data residency: all processing can stay in India. A data-residency view shows where each language model, recogniser, voice and data store processes data, and an India-only policy prevents selecting a provider outside India without a recorded approval.
• Privacy (DPDP Act, 2023): a caller’s data can be erased on request across call records, transcripts, recordings, analyses, metrics, callbacks and campaign lists, with the erasure itself audited; retention periods apply automatically; consent and recording notices are played where required.
• Logs and monitoring: audit and security events stream to the organisation’s SIEM over syslog or HTTPS; log retention is configurable with a minimum of 180 days.
• Outbound calling (TRAI): calling hours are limited to 09:00-21:00 IST; ordinary mobile or landline caller IDs are refused in favor of the 140 and 1600 series; every number is checked against a do-not-call list immediately before dialing; callers who ask not to be called again are added automatically; each campaign records its purpose, consent basis, DLT registration and auto-dialler notice before launch.
• Security testing: the platform is ready for VAPT, and its model layer is tested continuously with the evaluation framework’s security probes.

11. Operations and evidence#

AuraVoice treats evidence as a product feature:

• SLA reports from real call records: time to first audio (p50, p90, p95, by answer source), availability measured minute by minute, containment, abandonment, recognition, model and voice latencies, sentiment, cache hit rate.
• Machine-readable export of quality, latency, usage and SLA metrics (JSON for any period, and Prometheus format) with a dedicated monitoring token, so the organisation’s own monitoring, not a vendor dashboard, is the authoritative record.
• Failover log and drills, evaluation runs, speech-accuracy tests and the audit log, all retained.
• A live security checklist computed from the installation’s configuration.

12. Deployment#

AuraVoice runs as containers on a GPU server (an NVIDIA that runs in a complete self-hosted stack) or on Kubernetes, on premises, in an Indian cloud region, or in a hybrid of both, with self-hosted and hosted models mixed per role.

The installer prepares the NVIDIA drivers and container runtime, generates secrets and detects where the server runs; a GPU guard keeps the models on the GPU and repairs the service automatically; an uninstaller removes everything cleanly.

The platform is white-labelled per organisation.

---

About VedaCLOUD.AI#

VedaCLOUD.AI builds sovereign AI platforms for Indian government and enterprise. AuraVoice is designed, developed and owned by VedaCLOUD.AI.

To see AuraVoice answer your organisation’s questions in your languages, with your documents and on your telephony, contact connect@vedacloud.ai.