How Bhashini put the Red Fort speech into 22 languages in real time
On the 80th Independence Day, Prime Minister Modi's Red Fort address was translated live into all 22 scheduled Indian languages using Bhashini — the first time an AI pipeline has handled the speech at national scale. The system chains speech recognition, AI4Bharat's IndicTrans2 translation model and text-to-speech to deliver the address in each listener's mother tongue.
Manik Gupta
Founder and editor of DeepTech India. Manik writes about India's frontier technology ecosystem — AI, semiconductors, space, quantum, robotics and biotech — translating research and policy into clear, reliable reporting.
For 79 Independence Days, the Prime Minister's address from the ramparts of the Red Fort reached most Indians in a single language. On 15 August 2026 — the 80th Independence Day — it reached them in 22.
As Prime Minister Narendra Modi spoke in Hindi, his words were translated live into all 22 of India's scheduled languages using Bhashini, the open-source language-technology platform built under the Ministry of Electronics and Information Technology (MeitY). It was, by the government's account, the first time an Independence Day address had been run through an AI translation pipeline at national scale — a quiet but telling demonstration of how far India's home-grown language AI has come.
Three models, one sentence
What sounds like a single act of translation is, under the hood, a relay of three distinct AI systems working in sequence:
- Automatic Speech Recognition (ASR) converts the spoken Hindi into written text, in real time, as the Prime Minister speaks.
- Neural Machine Translation (NMT) takes that text and renders it into each target language. This stage is handled by IndicTrans2, a translation model developed by AI4Bharat, the research group based at IIT Madras that has become one of the country's most important open-source contributors to Indian-language AI.
- Text-to-Speech (TTS) turns the translated text back into natural-sounding audio, so a listener hears the address in their own language rather than merely reading a caption.
Chaining these three stages together — speech in, speech out, across 22 languages, with minimal lag — is a genuinely hard engineering problem. Each stage introduces potential errors that the next can compound, and doing it for languages with very different grammar, scripts and phonetics multiplies the difficulty. That the pipeline ran on a live, high-profile national event is a signal of growing confidence in the underlying stack.
Why 22 languages is the whole point
India's linguistic reality is that a message delivered only in Hindi or English bypasses a large share of the population. The 22 scheduled languages named in the Constitution are spoken by hundreds of millions of people, many of whom engage with government services in their mother tongue or not at all.
Bhashini's mandate has always been to attack this barrier as digital infrastructure rather than as a series of one-off translation projects. The platform is designed as a shared public resource — its models and datasets openly available — so that any department, startup or developer can plug multilingual capability into an app, a helpline or a public service. Deploying it on the Independence Day address is partly symbolic, but the symbolism has substance: it puts the government's own flagship AI to a visible, unforgiving public test.
AI as the day's connecting thread
The translation was not the only technology story of the 80th Independence Day. In his address, Modi set a target of training one crore — 10 million — young Indians in artificial-intelligence skills over the coming year, tying the pledge to the broader ambition of building a developed India by 2047. Officials also noted that, for the first time, the ceremonial equipment used at the Red Fort was indigenously built, and the flypast featured home-grown platforms including the Tejas light combat aircraft and the Prachand light combat helicopter.
Read together, the threads point in one direction: a state increasingly keen to showcase Indian-built technology — from the guns on the lawn to the AI in the audio feed — as an expression of self-reliance.
The reality check
It is worth keeping expectations grounded. A live translation of a set-piece speech is a favourable case: the setting is controlled, the speaker is clear, and the vocabulary, while wide-ranging, is not unpredictable. Real-world deployments — a farmer describing a crop problem, a citizen navigating a grievance line in a regional accent, a doctor dictating notes — are messier, and that is where speech-and-translation systems are still routinely tripped up.
The value of the Independence Day demonstration is therefore less about proving the technology is finished and more about showing it is trusted enough to put on the national stage. For AI4Bharat's IndicTrans2 and the Bhashini platform, that visibility is its own kind of validation, and a marker of how central Indian-language models have become to the country's public digital fabric.
If the ambition holds, the direction of travel is clear: a future in which speaking to — and being served by — the Indian state in one's own language is the default, not the exception.
Sources
Tags
More from Artificial Intelligence
Dubai Chambers and Nasscom Sign MoU to Open a Fast Lane for India's Agentic AI and Deep-Tech Firms
Dubai Chambers signed MoUs with Nasscom and FKCCI in Bengaluru on August 20 to build a formal channel for Indian agentic-AI and deep-tech firms to expand into Dubai, anchored to the emirate's own push to adopt agentic AI across its private sector.
Razorpay Builds Vulcan, an India-Made AI Foundation Model for Payments
Razorpay has unveiled Vulcan, which it calls India's first transformer-based AI foundation model built for payments — trained on nearly 3 trillion data points to raise success rates, cut fraud and smooth checkouts.