BHASHINI and IISc Bangaluru : Text-to-Speech Model for Sourashtra Language

1. At a Glance

• Digital India BHASHINI Division (DIBD) worked with IISc Bengaluru to build Text-to-Speech (TTS) models for Sourashtra. Sourashtra is a low-resource Indo-Aryan minority language spoken mainly in Tamil Nadu [1].
• It is a case study in AI for linguistic inclusion: it brings an under-represented language into digital services, education and accessibility tools [1].
• UPSC relevance: GS-I (Indian culture and diaspora communities), GS-II (inclusion and e-governance), GS-III (AI and IT). Prelims can test the agencies involved, the AI architecture, the dataset and the platform names [1].

2. Why in the News

• On 16 October 2026, PIB (Ministry of Electronics & IT) announced that DIBD and IISc Bengaluru had developed four Sourashtra TTS models [1].
• At the same time, the SPICOR TTS 3.0 Sourashtra Corpus (28+ hours of speech) was made available on BHASHINI's AIKOSH platform [1].

3. Background & Evolution

• Rationale: Languages with little digital data get left out of speech technology. This initiative aims to bring such languages into digital apps, language learning, accessibility tools, documentation and research [1].
• Parent body: BHASHINI is run by the Digital India BHASHINI Division (DIBD), under the Digital India Corporation (DIC), under MeitY. Its mandate is "AI-powered multilingual digital inclusion" [1].
• Milestones (from cited sources):
• April 2023: The Saurashtra-Tamil Sangamam was held in Gujarat from 17 to 26 April 2023, and the PM addressed the participants. It celebrated the historical link between Saurashtra (Gujarat) and the Sourashtra community of Tamil Nadu [3].
• January 2025: BHASHINI was used at Maha Kumbh 2025 to translate announcements, event schedules and safety guidelines into 11 Indian languages [2].

• 16 October 2026: The Sourashtra TTS models were released and the corpus was hosted on AIKOSH [1].

• Data source: The speech corpus was built at the SPIRE Lab, IISc Bengaluru [1].

4. Core Static Facts

Parameter	Fact
Developers	DIBD (under DIC, MeitY) + IISc Bengaluru [1]
Language	Sourashtra: minority Indo-Aryan language, mainly spoken in Tamil Nadu [1]
Speakers	About 2.48 lakh (Census 2011) [1]
AI architecture	VITS: Variational Inference with adversarial learning for end-to-end Text-to-Speech [1]
Number of models	4 = 2 voices (male, female) × 2 scripts (Tamil script and Sourashtra script) [1]
Training data	Male voice: about 10,000 utterances. Female voice: about 11,400 utterances [1]
Corpus	SPICOR TTS 3.0 Sourashtra Corpus: 28+ hours, gender-balanced, 1 male and 1 female voice artist [1]
Corpus content	Composed sentences plus sentences from literature. Transcribed in both Sourashtra and Tamil scripts [1]
Output audio	22,050 Hz [1]
Text input	Works directly on character-level text, so no separate phonemizer is needed [1]
Data platform	AIKOSH: BHASHINI's platform for language datasets and resources [1]
BHASHINI coverage	36 Indian text languages, 23 Indian voice languages, 35 international languages [1]
BHASHINI services	Translation, speech recognition, speech synthesis, transliteration [1]

5. Multi-Dimensional Analysis

Scientific / Technological

• VITS is an end-to-end TTS model. It learns to generate speech directly from text, so it does not need a separate acoustic model and vocoder pipeline [1].
• Working at the character level removes the need for a phonemizer. That matters for low-resource languages, which usually have no pronunciation dictionaries [1].
• A model for each script handles a language written in more than one script (Tamil and Sourashtra) [1].
• An open corpus on AIKOSH gives a foundation for follow-on research on Sourashtra language technology [1].

Social / Cultural

• Helps preserve a minority language with roughly 2.48 lakh speakers and limited digital presence [1].
• Gives the community access to digital content in its own language, plus speech-enabled learning tools for students [1].
• Accessibility: Helps users who rely on spoken digital content, such as people with visual impairments or low literacy [1].
• Cultural link: The Saurashtra-Tamil Sangamam (2023) highlighted the community's Gujarat–Tamil Nadu heritage [3].

Governance / Administrative

• Language AI can be built into government and public-facing digital services, which supports last-mile delivery of e-governance [1].
• The government–academia collaboration model (DIBD with IISc) can be repeated for other under-represented languages [1].
• BHASHINI has been used at large events, for example translation in 11 languages at Maha Kumbh 2025 [2].

Legal / Constitutional (these bullets are not from the whitelisted sources)

• Article 29(1): Any section of citizens with a distinct language or script has the right to conserve it. [static constitutional fact; not from the whitelist]
• Article 350A: Facilities for instruction in the mother tongue at the primary stage. Article 350B: Special Officer for Linguistic Minorities. [static constitutional fact; not from the whitelist]

Ethical

• Data representativeness: the corpus has only one male and one female voice artist [1]. That limits variety in dialect and speaker.
• Digital inclusion versus digital divide: the benefit depends on the community having devices and connectivity.

6. Recent Developments (last 12–18 months)

• 16 Oct 2026: Four Sourashtra TTS models (VITS) released by DIBD and IISc [1].
• 16 Oct 2026: SPICOR TTS 3.0 Sourashtra Corpus (28+ hours) made available on AIKOSH [1].
• As of Oct 2026: BHASHINI supports 36 text, 23 voice and 35 international languages [1].

7. Prelims Hooks

• Sourashtra TTS models were developed by DIBD (MeitY) with IISc Bengaluru (not IIT Madras) [1].
• DIBD sits under the Digital India Corporation, which is under MeitY [1].
• Sourashtra is an Indo-Aryan language, even though it is spoken mainly in Tamil Nadu (a Dravidian-majority state) [1].
• Sourashtra speakers: about 2.48 lakh (Census 2011) [1].
• Architecture used: VITS (Variational Inference with adversarial learning for end-to-end TTS) [1].
• 4 models: male and female voices, each in Tamil script and Sourashtra script [1].
• The corpus was built at the SPIRE Lab, IISc Bengaluru [1].
• SPICOR TTS 3.0 Sourashtra Corpus: 28+ hours, gender-balanced [1].
• The corpus is hosted on AIKOSH, BHASHINI's dataset platform [1].
• Output sampling rate: 22,050 Hz. The models use character-level input with no phonemizer [1].
• BHASHINI supports 36 Indian text languages, 23 Indian voice languages, 35 international languages [1].
• BHASHINI translated Maha Kumbh 2025 announcements into 11 Indian languages [2].
• The Saurashtra-Tamil Sangamam was held in Gujarat from 17 to 26 April 2023 [3].

8. Why a Voice Model Is Not Yet a Tool People Use

• The models speak in only two voices
• All 28+ hours of speech came from just one male and one female voice artist [1].
• So every Sourashtra sentence the models produce will sound like one of those two people.

• Speakers from other towns, older speakers and children may find the voice unfamiliar. That makes them less likely to trust or use it.

• The models learned from reading, not from everyday talk

• The training sentences were either written for the project or taken from literature [1].
• A model trained like this reads written text well. It has never heard how people actually talk at home or in the market.

• Most daily use, like a welfare message or a bus announcement, sounds like everyday speech, not like a book.

• The models only go one way: text to speech

• TTS (text-to-speech: the machine reads written text aloud) is what was released [1].
• The release does not mention ASR (automatic speech recognition: the machine understands what a person says) for Sourashtra [1].

• So a Sourashtra speaker can listen to a service but cannot yet speak to it. A person with low literacy still has to type or read to ask for anything.

• Two scripts split a small group of readers in two

• The models come in Tamil script and Sourashtra script [1].
• This is needed because the community writes in both. But it also means any app must first know which script its users read.

• With only about 2.48 lakh speakers [1], each script group is even smaller. Fewer users per version means less feedback to improve it.

• A model is not a service

• The release announces models and a dataset on AIKOSH [1]. It names no app, department or scheme that will use them, and gives no date.
• Until a ration office, school or helpline puts the voice in front of people, the benefit stays in the lab.

9. Why Small Languages Cost More to Bring Online

• Big languages get their data almost free. Small ones must make it from scratch
• For large languages, government already has huge piles of text. For example, the Lok Sabha Secretariat and MeitY agreed to build BHASHINI-based tools from Parliament's own records and debates [6].
• Sourashtra has no such pile. So IISc had to hire voice artists, write sentences and record them in a studio [1].

• That is why the corpus is only 28+ hours [1]. Every hour has to be paid for and recorded.

• The same cost will repeat hundreds of times

• India is among the countries with the largest number of endangered languages [4].
• UNESCO's atlas tracks about 2,500 endangered languages worldwide, ranked at five levels of danger [4].

• If each language needs its own studio project, India cannot cover them one by one. The method must become cheaper and repeatable.

• Is it worth it for 2.48 lakh people? Yes, but for a clear reason

• A fair question: why spend public money on a language with about 2.48 lakh speakers [1] when bigger languages still have gaps?
• The answer: once a language disappears, it cannot be brought back. UNESCO puts it simply: "when we lose a language, a community loses its unique vision" [5].
• But this answer holds only if the model is actually used. A model sitting unused on AIKOSH preserves nothing.

10. Who Should Do What Next

• DIBD and IISc should add more voices and a listening model
• Record more speakers of different ages and places, so the voice sounds like the whole community, not two people [1].

• Build ASR next, so people can speak to services and not only listen to them.

• DIBD should name a real user for the models

• The Sansad Bhashini MoU shows how this works: one institution (Lok Sabha) gives its data and uses the tools BHASHINI builds for it [6].

• A similar tie-up with Tamil Nadu departments working in Madurai and other Sourashtra areas would put the voice into actual services.

• MeitY should report use, not just launches

• Publish how many times the Sourashtra models are downloaded from AIKOSH and called by apps.

• Without this, nobody can tell if the project worked.

• The Ministry of Culture should link this to its language-preservation work

• The Ministry has said technology "may be the solution to a lot of our problems" in protecting languages [5].

• It can add Sourashtra recordings and models to its documentation work, so the data also serves as a record of the language.

• Use the UN decade as the frame

• UNESCO leads the International Decade of Indigenous Languages (2022–2032) and runs the World Atlas of Languages to track language status [5].
• India can report projects like this under that frame, and use the atlas to choose which language to do next, based on how endangered it is.

11. Anchors for Answers

• Data: About 2.48 lakh Sourashtra speakers (Census 2011); 28+ hours of speech from only 2 voice artists [1]
• Data: UNESCO's atlas tracks about 2,500 endangered languages worldwide; India is among the countries with the most [4]
• Report/Committee: UNESCO report on India's tribal and indigenous languages, launched 2022 [5]
• Law/Case: Article 29(1) (right to conserve a distinct language or script); Article 350A (primary education in the mother tongue); Article 350B (Special Officer for Linguistic Minorities)
• Comparison: Sansad Bhashini: Parliament's own records are used as ready-made data, which a small language like Sourashtra does not have [6]
• Scheme: International Decade of Indigenous Languages 2022–2032 (UNESCO) gives the global frame for language-technology projects [5]

12. Mains Relevance

• GS-I: Indian culture (languages, migrant communities, cultural synthesis).
• GS-II: Government policies for vulnerable sections and linguistic minorities; e-governance.
• GS-III: Science and technology (AI and IT developments, indigenisation of technology).
• Possible question stems: 1. "Artificial Intelligence can be a tool for linguistic preservation as much as for economic efficiency." Discuss with reference to BHASHINI's initiatives for low-resource Indian languages. 2. Examine the challenges in building AI language technologies for minority languages in India. How can government–academia partnerships address them? 3. "Digital inclusion in a multilingual society needs more than translation." Critically analyse.

13. Related Topics to Study Next

• BHASHINI / National Language Translation Mission: the parent mission, its structure and platforms.
• IndiaAI Mission and AIKOSH: India's AI dataset and compute ecosystem.
• Eighth Schedule and the demands for inclusion: status of non-scheduled languages like Sourashtra.
• Articles 29, 350A and 350B: constitutional safeguards for linguistic minorities.
• Saurashtra-Tamil Sangamam / Kashi-Tamil Sangamam: cultural-integration programmes linked to the same community.
• NEP 2020 and mother-tongue instruction: how language technology connects to multilingual education.
• UNESCO and endangered languages (International Decade of Indigenous Languages 2022–32): the global framework for language preservation.
• Digital India and the Rights of Persons with Disabilities Act, 2016: accessibility through speech technology.

14. NCERT Concepts to Revise

• Protection of interests of minorities (Polity): Article 29 gives any group with a distinct language or script the right to conserve it, the basis for this effort.
• Special Officer for Linguistic Minorities (Polity): Article 350B officer monitors safeguards for linguistic minorities like the Sourashtra community; it is a likely Prelims fact.
• Eighth Schedule (Polity): Sourashtra is a non-scheduled language; knowing the 22 listed languages explains its status and demands for recognition.
• e-Governance (Polity): Language AI like BHASHINI's speech models helps deliver government services to citizens in their own language.
• Digital India (Polity): BHASHINI is a Digital India initiative, run by the Digital India BHASHINI Division under MeitY.
• Digital divide (Geography): The note says benefits depend on devices and connectivity, so unequal access can limit who gains.

15. Common Errors / Trap Areas

• Language family: Sourashtra is Indo-Aryan, not Dravidian, even though its speakers live in Tamil Nadu [1].
• Partner institution: IISc Bengaluru (SPIRE Lab), not an IIT [1].
• Hierarchy: DIBD → DIC → MeitY. It is not under the Ministry of Education or the Ministry of Culture [1].
• Coverage numbers: Don't mix up 36 (text) with 23 (voice) Indian languages [1].
• Scripts: The models support both Tamil and Sourashtra scripts, not Devanagari [1].
• Spelling: "Sourashtra" (the language) vs "Saurashtra" (the region in Gujarat). Exam options may switch between the two.
• Not verified from the whitelist: BHASHINI's launch (reported as 4 July 2022, Digital India Week, Gandhinagar) and the Census grouping of Sourashtra under Gujarati. Check both against PIB or Census sources before relying on them.