Bhashini Offers Voice in 22 Indian Languages and Text in 36, and a Lab at IIT Madras Built the Open Translation Model for All 22 Scheduled Languages

Bhashini, the Government of India’s language platform, offered voice services in 22 languages and text services in 36 as of 9 February 2026, according to a backgrounder from the Press Information Bureau (PIB), the government’s press arm. The same document says the platform hosts more than 350 AI models and datasets and has completed over 4 billion language inferences, or 400 crore, each one a single run of a model on a request.

Open datasets and models for that platform come from AI4Bharat, a lab at the Indian Institute of Technology (IIT) Madras that serves as the mission’s data management unit. Its translation model, IndicTrans2, is described by its authors as the first to support all 22 languages in the Eighth Schedule of the Constitution.

A mission that began in July 2022

Bhashini is the working name of the National Language Translation Mission. The Ministry of Electronics and Information Technology (MeitY) runs it through the Digital India Bhashini Division, which the official site calls an independent division under the Digital India Corporation. The name expands to “BHASa INterface for India” in PIB’s August 2022 factsheet, which says the prime minister launched it at the inauguration of Digital India Week 2022 in Gandhinagar. Swarajya’s account puts the date at 4 July 2022 and reports an allocation of Rs 495.51 crore for three years.

The factsheet lists three aims. Citizens get access to the internet and digital services in Indian languages, including by voice. A crowdsourcing effort called Bhasha Daan collects language data from the public. And Bhashini acts as a coordinator that pulls contributions from government, industry, academia and start-ups into an open repository.

By February 2026 PIB described the services as translation, speech-to-text, text-to-speech, transliteration (writing a word in another script) and document understanding. The point, in its words, is that other platforms can add language and voice features without building them from scratch. The platform had passed 1.2 million downloads in PIB’s 10 February 2026 backgrounder, and the official site listed 1.4 million-plus mobile app downloads when tuput read it on 15 October 2026. The wider government AI programme, including its compute funding, is covered in tuput’s piece on India building AI in its own languages.

The lab at IIT Madras

AI4Bharat describes itself as a research lab at IIT Madras working on AI for Indian languages through open-source contributions. Its listed focus areas are transliteration, language understanding and generation, translation, speech recognition and speech synthesis. The funders named on its home page are MeitY, Rohini Nilekani Philanthropies, EkStep Foundation, C-DAC, Microsoft, Google and Yotta Infrastructure.

The page credits grants from MeitY for its data collection, which it calls part of the Data Management Unit of Bhashini. The IndicTrans2 paper says the same in its acknowledgements: the translators, reviewers and language experts on that project were supported by a grant from the Digital India Bhashini Mission to IIT Madras for that role.

In March 2023, IIT Madras Director V Kamakoti told Moneycontrol that researchers at the Nilekani Centre at AI4Bharat had a working text translator for nine languages, as part of Bhashini. The report adds that the centre began in 2022 with a Rs 36 crore grant from Nilekani Philanthropies.

IndicTrans2, the model for all 22 languages

India has 22 scheduled languages, listed in the Eighth Schedule, and the paper notes that about 97 percent of the population speaks one of them as a first language. For background on how those 22 were chosen, see tuput’s explainer on the Eighth Schedule.

The IndicTrans2 paper, by 14 authors led by Jay Gala and Pranjal Chitale, was posted to arXiv on 25 May 2023 and published in Transactions on Machine Learning Research in December 2023. Its abstract says that before this work there was no parallel training data covering all 22 languages (the same sentences in two languages, which a model learns from), no benchmark covering all of them, and no translation model supporting all of them.

The team built three things. The Bharat Parallel Corpus Collection (BPCC) holds about 230 million sentence pairs, of which 126 million were newly added, including 644,000 manually translated pairs. IN22 is a test set the authors call the first benchmark with the same content in all 22 languages. IndicTrans2 itself has 1.1 billion parameters, with distilled versions of about 211 million, and a second version translates directly between Indian languages without passing through English, covering 462 directions.

On quality, the authors report that IndicTrans2 performs at par with the open NLLB model from Meta and with commercial systems from Google and Microsoft, using automatic scores called BLEU, chrF++ and COMET. These are the authors’ own evaluations. The paper also says the Workshop on Machine Translation, a long-running shared task, had covered only 4 Indian languages in total over 16 years.

Models and data were released under open licences. The GitHub repository lists model checkpoints under the MIT licence, and its release notes record a long-context update on 18 January 2025. India’s AI repository, AIKosh, also lists an IndicTrans2 entry with the Digital India Bhashini Division as source organisation. None of the pages tuput opened says which model runs inside a particular Bhashini service, so this article does not claim that one.

The text and speech data underneath

IndicTrans2 draws on earlier releases from the same group. The first was IndicNLPSuite, published in Findings of EMNLP in November 2020. It offered monolingual text of 8.8 billion tokens across 11 languages and Indian English, drawn mostly from news crawls, and pre-trained language models built on the compact ALBERT design.

The ACL 2023 follow-up, by Sumanth Doddapaneni and colleagues, raised the corpus (IndicCorp) to 20.9 billion tokens in 24 languages, about 2.3 times the earlier size. It also introduced IndicXTREME, a benchmark of nine tasks across 20 languages with 105 evaluation sets, 52 of them new, and IndicBERT v2, a language model that the authors say improves on a strong baseline by 2 points on average.

For speech, the Kathbath corpus in the IndicSUPERB paper holds 1,684 hours of labelled speech in 12 languages from 1,218 contributors in 203 districts. IndicVoices, posted in March 2024, has 7,348 hours from 16,237 speakers across 145 districts and 22 languages, of which 1,639 hours were transcribed. Only 9 percent of it is read text, against 74 percent extempore speech and 17 percent conversation.

Where it is being used

PIB’s February 2026 backgrounder gives two events. At the Kashi Tamil Sangamam in Varanasi, the prime minister’s Hindi speech was translated into Tamil in real time through Bhashini. At the Maha Kumbh 2025, a Bhashini-powered chatbot called Kumbh Sah’AI’yak gave pilgrims navigation and event information in 11 languages and ran a lost-and-found feature with live text and voice translation. The official site also quotes the prime minister saying that at the Supreme Court’s Diamond Jubilee celebrations in 2024 his speech was translated into English in real time.

Parliament signed up on 18 March 2025. The Lok Sabha Secretariat and MeitY signed an agreement for Sansad Bhashini, in which Parliament supplies its archive of debates to fine-tune the tools and Bhashini supplies translation. The PIB release lists four planned uses: translation of past debate documents into regional languages, a chatbot for the Parliament website, speech-to-text in Indian languages, and speech-to-speech conversion with real-time transcription. The Bhashini home page now carries a Sansad Bhashini banner describing translation, transcription and summarisation for Parliament.

On the railways, the Digital India Bhashini Division and the Centre for Railway Information Systems (CRIS) signed an agreement on 9 June 2025 to bring text, speech and character-reading tools in 22 languages to systems such as the National Train Enquiry System and RailMadad, the complaint portal. All India Radio reported that workshops and pilots were due to begin in the following months. Bhashini’s Our Impact page separately lists the enquiry system as a service using text and speech recognition, along with the government’s CPGRAMS grievance portal, which it describes as offering text in 22 languages and voice in 16.

In the courts, the Supreme Court translates judgments into 18 Indian languages with its SUVAS tool, built with the National Informatics Centre, and hosts them on the e-SCR portal, according to PIB’s 11 February 2026 backgrounder. The backgrounder says translations are reviewed within judicial frameworks and does not say which language engine SUVAS uses. Tuput’s report on AI in Indian courts has the full list of court tools and the documents behind them.

What the published numbers leave open

The IndicTrans2 paper is direct about where work remains. It says the models generalise only to a limited degree for Manipuri (Meitei), Santali and Sindhi in Devanagari script, and that a performance gap remains between languages with plenty of data and those with little. The authors cite scarce human-annotated data and little web content as the reasons.

Counts of what the platform supports differ by service and by date. PIB’s February 2026 backgrounder gives 22 voice languages and 36 text languages, while the official site lists voice in 16 languages for one grievance system. The pages opened for this article give no accuracy figures for deployed services, and the railway and Parliament agreements describe planned work. Independent comparisons of speech systems in Indian languages are collected in tuput’s roundup of voice and translation models.

AI4Bharat’s own pages list a goal of 15,000 hours of transcribed speech from over 400 districts, which is a target and not a completed total.