India is home to around 1.4 billion people. But access to information here is not evenly distributed — it is shaped by four deep, overlapping divides.
- Literacy divide — nearly one in five Indian adults cannot read or write. Rural literacy hovers around 77.5%, and the gender gap still exceeds 12 percentage points. For millions, text is not an interface — it's a barrier.
- Financial divide — India is one of the top contributors to the world's unbanked adults. Large segments remain outside formal systems, which directly affects how they access services and support.
- Digital divide — around 350 million Indians still use feature phones. No apps, no mobile internet, no AI assistants. Smartphone penetration is under 50%.
- Demographic divide — almost half of Indians above the age of 60 are illiterate. Elderly citizens, rural households, and marginalised women often sit at the intersection of all four divides — excluded not by intent, but by design.
Yet almost every AI product today assumes literacy, internet, and a smartphone — serving the people who are already connected, not the ones who are excluded. Ironically, the people excluded by this design are the ones who need AI support the most.
"But there is one constant across 1.4 billion Indians. Voice. Literate or not. Connected or not. Urban or rural. Every individual speaks. Every individual listens — in their own language, in their own dialect."
Telephony integration is stable. The models are there, the pipelines exist, the demos are impressive. So the real question is no longer: 'Does Voice AI work?' The real question is harder — can it run as public infrastructure at population scale, on a government budget, without routing every citizen's voice through someone else's data center?
At the India AI Impact Summit, VoicERA was unveiled as India's first open-source Voice AI stack — a breakthrough in how voice systems are built and deployed, on cloud or on-prem, with complete ownership and control.
01 — The landscape is ready. The infra model isn't.
India's voice ecosystem has reached real technical maturity — and that deserves to be stated clearly. AI4Bharat's Conformer-based Indic STT now delivers production-grade accuracy. Modern Indic TTS systems generate speech that sounds natural. Open-source LLMs like Llama handle complex reasoning reliably. Real-time WebSocket telephony pipelines operate with stability and low latency.
The core components are no longer experimental.
But the way these pieces get deployed hasn't caught up with the ambition.
Most voice deployments today follow a familiar pattern: centralized cloud hosting, dependence on H100-class GPUs, ₹2.7–₹6 per minute recurring costs, audio data residing outside the originating institution, and significant friction when attempting to switch vendors. For a proof of concept, that model is acceptable. For a limited pilot, it's workable. But as long-term public infrastructure serving millions of citizens at scale, it collapses.
02 — Three numbers that change the conversation
The first is cost, and it compounds fast
At first ₹2.7–₹6 per minute sounds harmless. But scale changes the equation completely. India has roughly 40 million farmers. Take just one state — Madhya Pradesh — with about 3 million farmers. Even if each farmer makes one call per week, that's 156 million calls per year.
- ₹42.1 Cr – ₹93.6 Cr
- annual cost for a single state, single use case
- 156M
- calls per year (1 call/week from 3M farmers)
- ₹0.28
- per 2.5-min call on Voice-in-a-Box (amortised)
Public systems are not SaaS experiments with variable demand. They are continuous infrastructure. Recurring per-minute billing turns voice into structural OPEX that public institutions are not designed to carry.
The second is data, and this is fundamentally a policy question
Voice is not just audio. It carries identity cues, dialect patterns, tone, intent, and contextual information that can easily qualify as sensitive personal data. It often contains explicit PII — names, phone numbers, land records, grievance details — spoken naturally during conversation.
Under India's DPDP framework, this raises critical questions: Where is the audio stored? Where is it processed? Can retention and deletion policies be enforced locally? Who owns the transcripts and logs? What happens to that data when a vendor contract ends or pricing terms change?
The third is hardware, and the assumptions baked into it
Most voice stacks are designed with hyperscale GPU clusters in mind. District health centres, state agricultural departments, and taluk-level offices do not operate like hyperscale cloud. If Voice AI requires enterprise GPU clusters to function, it will never become sovereign public infrastructure — it will remain a pilot at the top of the system, never reaching the bottom.
Voice inference does not inherently require H100-class hardware. Its sequential, low-batch nature often makes mid-tier or consumer-grade GPUs more cost-efficient for real-time deployments.
03 — What VoicERA actually is
VoicERA is not a chatbot. It is a complete, open-source, production-grade Voice AI execution hardware and software stack — delivered as an open Digital Public Good, deployed on BHASHINI infrastructure or in your own data center, and designed to run in environments that look nothing like a hyperscale data centre.
The stack covers the full pipeline: telephony intake, STT, LLM reasoning, TTS output — with modular API switching at each layer, so individual components can be replaced as better models emerge without rebuilding the plumbing. It includes an Agent Builder platform for configuring use-case-specific behaviour, an observability dashboard for monitoring latency, concurrency and call success rates in real time, and governance controls for managing access, logging and retention policies.
04 — Voice-in-a-Box: the part that changes procurement math
Voice-in-a-Box means the entire voice pipeline — STT, LLM, TTS, orchestration, session management, logging, and governance — runs locally. On hardware the institution owns. In infrastructure the institution controls. With audio that never leaves.
The hardware philosophy is deliberately pragmatic. Instead of assuming hyperscale H100-class clusters, Voice-in-a-Box is engineered to operate on consumer-grade or workstation-class GPUs.
- A single 16GB GPU can handle 4–6 concurrent calls depending on model mix.
- Higher concurrency (50–100) uses larger multi-GPU systems (80GB-class).
- JOHNAIC-class appliance builds ship as modular off-the-shelf nodes.
- Fully containerised, portable, hardware-agnostic. Cloud remains available — but it is no longer the dependency.
| Model | Cost structure |
|---|---|
| Cloud voice stack | ₹2.7–₹6 per minute, recurring, indefinitely |
| Voice-in-a-Box | One-time hardware investment; effective per-minute cost approaches ₹1 at sustained high volumes. |
OPEX → CAPEX
A structural shift from perpetual running cost to a predictable infrastructure investment that depreciates in a budget line someone can actually plan around.
06 — VoicERA builds on the power of DPGs
A Digital Public Good (DPG) is open-source infrastructure designed for public use, governed transparently, and built to prevent vendor lock-in at the system layer. India has already seen the power of DPI — Aadhaar built a universal identity layer, UPI transformed payments into instant interoperable rails, DIKSHA created a shared digital backbone for education. VoicERA can play a similar role: a sovereign, interoperable conversational layer that allows every institution to deliver services through speech — not just to the digitally connected, but to everyone.
07 — The engineering underneath
For the technically inclined: this is not stitched glue code. The stack includes real-time WebSocket telephony integration, session orchestration with token budgeting, GPU scheduling and load balancing, concurrency scaling from 25–50 calls in Phase 1 to 50–100 in production targets, noise cancellation and audio normalisation, the Agent Builder for prompting and sandbox testing, and a full observability layer.
08 — The questions that come next
- How does the stack behave during seasonal call spikes — the Kharif planting window in Maharashtra, for example?
- How are distributed appliance nodes monitored across districts with variable connectivity?
- What's the rollout protocol for model updates on on-prem deployments?
- How do we expand from 22 scheduled languages toward 700 dialects?
- How do we sustain low latency and high concurrency in production, with real-world noise and network variability?
- How do we ensure continued improvement of the voice models — and reduce or eliminate LLM hallucinations for safe use?
"Voice infrastructure behaves more like a telecom network than a SaaS product. The failure modes are different. The tolerance for degraded service is essentially zero when the person on the other end is a farmer asking about crop disease."
The real bet
Voice AI is powerful. Cloud-dependent voice is expensive, centralised, and fragile at public scale. Voice-in-a-Box makes it sovereign, affordable, deployable, and portable — to the last district, the last taluk, the last farmer who speaks a dialect that no enterprise product has ever been optimised for.
VoicERA is the infrastructure layer — the thing underneath the thing — for how public systems will speak to citizens and listen to them at scale. That is a much more important ambition than launching another bot.