01.A speech AI push from Mysuru
On 27 October, the Central Institute of Indian Languages in Mysuru launched Anuvani, an open-source ecosystem of speech AI models built for India's languages. The institute works under the Department of Higher Education in the Ministry of Education. The launch begins with two models. Anuvani-LID identifies which language is being spoken. Anuvani-Diarization identifies who is speaking and works out when each person speaks.
Both models support all 22 constitutionally scheduled Indian languages and Indian English. CIIL is releasing them together with peer-reviewed research papers, open-source model weights, on-device mobile toolkits and production-grade NVIDIA Triton Inference Server kits. The stated aim is that developers, researchers, startups and public institutions can use and deploy the technology without proprietary licensing barriers.
02.Why Indian speech is hard for machines
CIIL says the models are built for conditions that general-purpose speech systems often struggle with in India. Conversations here frequently move between closely related languages, carry very different regional accents, include speakers talking over each other, and are recorded in noisy surroundings. A system trained mostly on clean English audio tends to fail on all four. Speech also mixes languages mid-sentence in ordinary Indian usage, which breaks systems that expect one clean language label per recording.
The diarisation model addresses the hardest part of that list. Working out who spoke when is easy in a studio interview and much harder in a phone call, a meeting room, a large hall or a broadcast, especially when speakers change quickly or overlap. Getting this step right is what makes automatic notes, call analytics and searchable recordings possible at all.
03.What the numbers say
CIIL academic secretary Tariq Khan gave details of both models. Anuvani-LID uses a Conformer architecture with 115.6 million parameters and was trained on 54,643 hours of Indian speech, covering more than 31 million utterances. Anuvani-Diarization is smaller at 76.1 million parameters and is designed for phone calls, meetings, recordings made in large rooms and broadcast audio.
The diarisation model was tested on IndicDiarBench, a CIIL benchmark of 1,164 recordings totalling 107.7 hours across all 22 scheduled languages. In a zero-shot evaluation, meaning the model was not tuned on that benchmark beforehand, it recorded the lowest Diarization Error Rate among the 18 systems tested. The scoring standard was strict and included overlapping speech, the situation most systems handle worst. Diarisation Error Rate measures how much speaking time a system attributes to the wrong speaker or misses altogether, so a lower figure means cleaner separation.
04.What it means for students, developers and small firms
The practical difference is licensing and effort. A team building a voice product in Kannada, Bengali or Marathi normally either pays for a commercial speech stack or spends months assembling language detection and speaker separation from scratch. Open weights plus ready deployment kits change the starting point. The on-device toolkit matters for apps that must work on a phone without a reliable connection, and the server kit matters for organisations running on cloud infrastructure. That does not remove the engineering work. A team still has to collect and label audio from its own users, because accents, names and local vocabulary shift from district to district, and it still has to build a product around the model.
For students and researchers, the peer-reviewed papers and published weights are a usable base to fine-tune and benchmark against. For small businesses, the obvious uses are call-centre recordings, helpline logs, lecture capture, meeting notes and accessibility tools. For public institutions, CIIL says the intent is deployment at any scale, from cloud clusters to offline smartphones.
05.A shared resource, with limits
Basavaraja Kodagunti, director of CIIL, framed the launch as continuity rather than a break. He said the institute has worked for over five decades to preserve, document and advance India's rich linguistic heritage, and that Anuvani brings that deep linguistic foundation into the frontier AI era. He added that releasing the papers, weights and toolkits would help keep Indian languages a shared public resource and let citizens, developers and institutions take part in the digital future in their mother tongues.
It is worth being precise about scope. This first release covers two tasks, language identification and speaker separation. It is not a full speech recognition or translation stack, and the wider ecosystem will only be as useful as the models CIIL adds next. The real test will be adoption: whether startups, colleges and government departments actually build on these weights over the coming year. Open weights also make claims easier to check, since anyone can run the same benchmark and compare the result.
Frequently asked questions
What is Anuvani?
It is an open-source ecosystem of speech AI models launched by the Central Institute of Indian Languages, Mysuru, on 27 October 2026. Its first two models identify the language being spoken and separate who spoke when, across all 22 scheduled Indian languages and Indian English.
Is it free to use?
CIIL has released the models with open-source weights, peer-reviewed papers, on-device mobile toolkits and server deployment kits, without proprietary licensing barriers. Developers, researchers, startups and public institutions can deploy them without a commercial licence.
Can it turn speech into text?
Not in this first release. The two models handle language identification and speaker diarisation, which are the steps that usually come before transcription. A full speech-to-text model is not part of what CIIL has announced so far.