Voice AI: India’s next digital leap must include every Indian’s language

For decades, technology has dictated the terms to us: type, click, tap, swipe. What if we could turn that around and simply “dictate” to technology? 

Recently, I was filling out a form for my family of five, working through field after field of questions. Halfway through, I thought: why am I typing all of this? Why can’t I just speak? 

Imagine saying what you need, in your own language and having a digital system understand and help you get it done. For India’s Digital Public Infrastructure (DPI), this means bringing public services to citizen’s fingertips, regardless of language, literacy or digital proficiency. That is the promise of Voice AI: making universal digital access an everyday reality. 

DPI and Voice AI: Making public services easier to use 

One needs to think of India’s DPI ecosystem in three layers: foundational systems such as Aadhaar, UPI, and DigiLocker; sectoral platforms such as e-NAM, ONDC, and the Ayushman Bharat Digital Mission; and the apps and portals through which citizens access services, such as UMANG. 

Voice can improve that last layer of apps and portals, helping new users navigate services and saving time for those comfortable with technology. As both an inclusion tool and a convenience layer, Voice AI could make digital services more intuitive for everyone, helping more people get started and keep coming back. 

Think of a buyer in India’s hinterland searching, comparing products and placing an order on ONDC simply by speaking. Or a citizen asking DigiLocker for a licence or certificate and receiving spoken guidance to find it. With over 20 crore on ONDC and more than 72 crore registered users on DigiLocker, even modest improvements in ease of use could have a substantial impact. We estimate that voice-led access could increase activation among underserved users by 15–25% and improve completion or repeat engagement by 25-30%. 

The Next Frontier: Voice AI helping citizens complete tasks 

Voice AI is moving beyond answering questions to carrying out requests. Connected to authorised systems, an agent could identify the steps involved, retrieve relevant information and help complete a task or transaction. 

It could help a citizen submit a government claim, support a student learning in their preferred language or help a banking customer resolve a problem without navigating repeated “press 1, press 2” menus. 

States could build VoiceAI-DPI: voice-powered public infrastructure drawing on their own administrative records across land, agriculture, health and welfare. The apex advantage lies in connecting it all: Voice AI for access. State data for intelligence. Government systems for action. A farmer could ask about a scheme in their own language, understand eligibility and receive help applying. 

Over time, connected records could help states identify when someone may qualify for support and offer timely guidance. Privacy must be built into these services from the start, in line with India’s Digital Personal Data Protection Act, 2023, and applicable rules. Citizens should be told, in their own language, what is recorded, which personal records are accessed and why. 

What needs to work at scale 

Making Voice AI work requires departments to connect their systems, keep records current and agree on how requests move between them. A conversation cannot resolve an application stuck between offices. 

The administrative work matters as much as the interface. Pilots should establish whether this works before services expand statewide. 

Language accuracy is fundamental. India has 22 official languages and many dialects. People switch languages within sentences, use local expressions and speak with different accents. Add the risk of Large Language Model (LLM) hallucinations. A system must handle those variations without confidently turning a misunderstood phrase into an incorrect action. Better training data, local vocabulary and testing with intended users are essential. When a request is unclear, the system should ask. Larger models alone cannot substitute for understanding the conditions in which people will use a service. 

Speed matters too. Even accurate Voice AI can deliver a poor experience if the response feels slow. Leading platforms are targeting end-to-end latency of roughly 500 to 800 milliseconds, but the underlying technology architecture can significantly influence whether an interaction feels natural or frustrating. 

Affordability requires similar scrutiny. The cost includes model use, integration, maintenance and human assistance. Cost per successfully completed task is a more useful measure than a low price per minute. Repeated attempts quickly undermine both savings and confidence. 

India has shown it can build public digital infrastructure at scale. Its promise comes to life when people can use it with confidence and independence, in the language they call their own. 

Voice AI can help make that possible, provided it understands India’s many languages, protects people’s information and helps them complete what they came to do. Behind every request is a person trying to get something done. Being understood is the beginning; getting the help they need is the goal. A digital nation fulfils its promise when every person feels it was built for them. 

Voice AI can convert India’s linguistic diversity from a barrier to an opportunity as digital experiences become not technology-centric but citizen-centric. 

(The author is Chief Strategy Officer, Deloitte South Asia. Views are personal.)