IISc's SraVaani brings open-source speech AI to 65 Indian languages

Most voice assistants have a blind spot. Ask them to transcribe Hindi, Tamil or Bengali and they cope; switch to Tulu, Kokborok or Bundeli and they fall silent. For tens of millions of Indians, that gap is the difference between being able to speak to a machine and being locked out of the digital services increasingly built around voice.

On 13 August 2026, the Indian Institute of Science (IISc), Bengaluru, released a model aimed squarely at that gap. Called SraVaani, it is an open-source multilingual automatic speech recognition (ASR) system — software that turns spoken words into written text — covering 65 Indian languages and dialects. The model was built by IISc's SPIRE Lab in collaboration with the AI and Robotics Technology Park (ARTPARK), with support from Google.

What SraVaani actually does

At its core, SraVaani is a speech-to-text engine, the first stage in almost every voice-driven application: dictation tools, call-centre automation, voice search, real-time captioning and the front end of conversational assistants. What sets it apart is breadth. The model spans:

20 of India's 22 scheduled languages, the constitutionally recognised tongues that anchor official communication.
45 additional regional languages and dialects, many of which have had little or no prior voice-AI support — among them Garo, Angika, Chakma, Kokborok, Tulu, Bundeli and Bajjika.
10 scripts, reflecting the visual diversity of Indian writing systems.
Automatic language identification, so the system can recognise which language is being spoken without the user having to specify it in advance — a practical necessity in a country where speakers routinely switch between tongues.

That last feature matters more than it sounds. In a multilingual setting, a rigid model that assumes one input language will stumble the moment a speaker code-switches. Automatic identification lets a single deployment serve a linguistically mixed population without manual configuration.

Why the long tail is the hard part

Building strong ASR for a well-resourced language is largely a question of data: the more transcribed audio a model sees, the better it gets. Global tech firms have poured resources into English, Mandarin and a handful of European languages precisely because the training material is abundant.

India's smaller languages sit at the opposite end. Speech data is scarce, transcriptions rarer still, and commercial incentives thin. The result is a self-reinforcing neglect: because there is little data, models perform poorly; because models perform poorly, the languages stay off digital platforms; and because they stay off, no new data accumulates. Reaching 45 regional languages and dialects means confronting exactly this scarcity — and doing so is the whole point of a publicly funded research effort rather than a purely commercial one.

Open source as a deliberate choice

Releasing SraVaani as open source is a strategic decision, not an afterthought. It means startups, government departments, universities and independent developers can build on the model without licensing an equivalent from a foreign provider — and can adapt it for their own domains, whether that is healthcare, agricultural advisory services or public-grievance systems.

For a country pursuing technological self-reliance, a home-grown, freely available speech stack reduces dependence on proprietary APIs whose pricing, availability and language coverage are set elsewhere. It also invites the very community that speaks these languages to contribute data and corrections, which is arguably the only sustainable way to keep improving the long tail.

Where it sits in India's language-AI wave

SraVaani lands amid a broader surge of Indian-language AI. Startups such as Sarvam AI have released voice-optimised foundation models, government platforms like Bhashini are stitching speech and translation tools into public services, and academic consortia are training model families across all 22 scheduled languages. What distinguishes SraVaani is its explicit tilt toward the underserved end of the spectrum — the dialects that rarely make it into a commercial roadmap.

The practical payoff, if the model performs in the field as advertised, is inclusion: farmers, patients and citizens who speak a regional dialect could dictate a query, navigate a voice menu or receive a spoken response in their mother tongue rather than being forced into Hindi or English.

What to watch next

An open-source release is a starting line, not a finish. The real tests are accuracy across noisy, real-world audio; how quickly developers adopt it into shipping products; and whether the community contributes the additional data needed to lift performance for the smallest languages. Voice AI also tends to expose edge cases — accents, background noise, mixed-language speech — that only surface at scale.

For now, SraVaani is a substantive addition to India's growing stack of sovereign language technology, and a reminder that the hardest and most valuable work in Indian AI may lie not in the languages everyone already serves, but in the dozens that almost no one does.

Sources

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.

4 min read

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.