Examine the challenges in building AI language technologies for minority languages in India. How can government–academia partnerships address them?

India is among the countries with the most endangered languages [2]. Yet AI language tools need large amounts of digital data, and minority languages rarely have it. The Text-to-Speech (TTS) models for Sourashtra, built by the Digital India BHASHINI Division (DIBD) and IISc Bengaluru, show both what these tools promise and how hard they are to build. Sourashtra is an Indo-Aryan language with about 2.48 lakh speakers, mostly in Tamil Nadu [1].

Challenges

• Data scarcity: Big languages already have large bodies of text. For example, Parliament's records feed Sansad Bhashini [4]. Sourashtra needed studio-recorded speech, which produced only 28+ hours [1].
• Narrow representation: The corpus uses only one male and one female voice artist [1]. It misses differences in dialect, age and speaker.
• Bookish, one-way data: The training sentences were composed or taken from literature, and only TTS has been released [1]. There is no speech recognition yet, so users with low literacy can listen but cannot speak to services.
• Multiple scripts: Tamil and Sourashtra scripts each need separate models [1], which splits an already small user base.
• Lab-to-service gap: The release names no app or scheme that will use the models [1].
• Scale and cost: UNESCO's atlas tracks about 2,500 endangered languages worldwide [2]. One-off studio projects cannot cover them all.

Role of Government–Academia Partnerships

• Combining strengths: DIBD provides the platform. IISc's SPIRE Lab built the corpus and the VITS models [1].
• Low-cost, reusable methods: The models read text character by character and need no phonemizer, the pronunciation tool most small languages lack [1]. This makes them a template that can be reused for other languages.
• Open data commons: The SPICOR corpus is hosted on AIKOSH, so others can build on it [1].
• Assured users: In the Sansad Bhashini MoU, the Lok Sabha supplies data and uses the tools built from it [4]. Similar tie-ups with Tamil Nadu departments could put these voices into real services.
• Wider corpora and documentation: Academic fieldwork can add more speakers and speech-recognition models. It can also link to the Ministry of Culture's language-preservation work, which sees technology as part of the solution [3].
• Outcome tracking: MeitY could publish how often the AIKOSH models are downloaded and used.

Building AI for minority languages is hard because of data, representation and deployment gaps. Partnerships can make the effort repeatable and accountable. Scaling them under the UN International Decade of Indigenous Languages (2022–32) [3] would help fulfil Article 29(1)'s promise that every community can conserve its language and script.