![CIIL Anuvani: Open Speech AI for 22 Languages [2026]](https://shattered.io/wp-content/uploads/2026/10/ciil-anuvani-open-source-speech-ai-22-indian-languages-2026-gen.webp)
India’s national language institute has put speech AI for the country’s 22 scheduled languages into open hands. The Central Institute of Indian Languages (CIIL) launched Anuvani, an open-source speech-AI ecosystem, and its first two public models, Anuvani-LID and Anuvani-Diarization, cover all 22 scheduled languages plus Indian English, according to The Hindu and the Hugging Face listings.
Anuvani-Diarization answers one question in multi-speaker audio: who spoke when. Release 1.0.0 ships a Python package called anuvani_diar, a Triton Inference Server and FastAPI serving kit, and the original Python reference implementation. CIIL also published Indic DiarBench, a dataset of 1,164 recordings that totals 107.7 hours across 22 Indian languages.
This release differs from the transcription models that dominate recent speech headlines. Nvidia’s Nemotron 3.5 Transcribe, which covers 40 languages, targets turning audio into text. CIIL Anuvani targets two plumbing tasks that sit around transcription, and it builds them around India’s own language list. Below we cover what CIIL confirmed, what it left open, and what the launch means for developers and companies that build voice products for Indian users.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What CIIL Released With Anuvani #
CIIL describes Anuvani as an ecosystem, not a single model. That word matters. An ecosystem implies a family of releases that share data, tooling and licensing conventions, and the Hugging Face listing groups the work under a collection titled “Anuvani Collection.” Two members of that collection are public today: Anuvani-LID and Anuvani-Diarization. Both are identified as open-weight speech models.
The Hindu reported the launch and framed it around language coverage. The models handle all 22 scheduled languages of India and Indian English. That list includes widely spoken languages and several with far smaller digital footprints, which is the core argument for a national institute building this layer rather than leaving it to commercial vendors that chase the biggest markets first.
The release package for the diarization model is unusually complete for a first public version. It bundles installable code, a production-style serving kit and the research reference implementation. Many academic releases ship only weights and a notebook. A team that wants to test Anuvani-Diarization against its own call recordings can read the package, run the server and compare it to its current pipeline within a day.
Several details remain open. The sources we reviewed do not confirm the exact launch date, the launch venue, the development team or individual creators, or the pricing and commercial-use terms of the wider ecosystem. They also do not confirm parameter counts or accuracy scores beyond the dataset details covered below. We treat all of those as unknown and flag them again in the confirmed-versus-unconfirmed table later in this article.
Anuvani-LID: Language Identification Across 22 Languages and Indian English #
The “LID” in Anuvani-LID points to spoken language identification, the task of deciding which language a clip contains before anything else happens to it. The sources confirm that the model is open-weight and that it covers the 22 scheduled languages plus Indian English. They do not give a parameter count, a training-data size or a benchmark score, so we do not repeat any figures floating around elsewhere.
Language identification looks like a minor task until a product has to serve a country where a single phone call can switch languages mid-sentence. A support line in Bengaluru might hear Kannada, English and Hindi inside one minute. A recognizer tuned for one language will fail loudly or, worse, fail quietly with plausible-looking nonsense. A reliable first-stage identifier lets a pipeline pick the right recognizer, translation route or reply language. That is our analysis of the typical design, not a claim from CIIL’s release notes.
Including Indian English as its own category is a practical choice. English spoken in India has accents, rhythms and borrowed vocabulary that differ from the English that general-purpose systems train on most heavily. Treating it as a first-class label gives product teams a cleaner routing decision than forcing it into a generic English bucket.
Open weights also change who can inspect the model. A closed language-ID service gives a label and sometimes a confidence value. Open weights let a research group probe where the model confuses closely related languages, which is exactly where Indian language identification gets difficult. Expect independent audits of Anuvani-LID once outside researchers start testing it.
Anuvani-Diarization: Who Spoke When in Multi-Speaker Audio #
Speaker diarization splits a recording into segments and labels each one by speaker. The Anuvani-Diarization model is designed to identify who spoke when in multi-speaker audio. In practice, that output is what turns a raw meeting recording into a transcript with speaker turns, or a call recording into separate agent and customer tracks.
What Release 1.0.0 ships
The model card lists three deliverables. First comes the anuvani_diar Python package, which wraps the pipeline for direct use. Second comes a Triton Inference Server and FastAPI serving kit, which gives operations teams a path to run the model as a service. Third comes the original Python reference implementation, which lets researchers see the logic behind the packaged version.
The serving kit deserves attention. Triton is a common choice for teams that already run GPU inference at scale, and FastAPI is a common choice for small teams that want an HTTP endpoint quickly. Offering both lowers the cost of a trial for both kinds of team.
The upstream networks inside the pipeline
Anuvani does not claim to reinvent every layer. The model card identifies the segmentation network as DiariZen’s WavLM-Large and Conformer network, using the checkpoint named BUT-FIT/diarizen-wavlm-large-s80-md-v2. It identifies the voiceprint network as pyannote/wespeaker-voxceleb-resnet34-LM. The weights CIIL lists include segmentation, PLDA and voiceprints.
Building on established components is the sensible route. WavLM-Large style encoders and Conformer segmentation have a track record, and the real contribution here is the adaptation and evaluation on Indian languages. A reader should judge Anuvani-Diarization on how well that adaptation holds up across the 22 languages, not on architectural novelty. CIIL’s sources do not publish an error rate that we can verify, so that judgment has to wait for independent tests on Indic DiarBench.
Indic DiarBench: 1,164 Recordings and 107.7 Hours #
A model without a shared test set invites unfalsifiable claims. Indic DiarBench is CIIL’s answer for diarization. The data contains 1,164 recordings totaling 107.7 hours, spans 22 Indian languages, and covers three recording conditions. The dataset revision listed is 92877bad, which gives researchers a fixed reference point so that results from different labs compare cleanly.
Some simple arithmetic puts those numbers in context. Dividing 107.7 hours by 1,164 recordings gives an average of about 5.6 minutes per recording. That is short enough to suggest conversational clips rather than long lectures or parliamentary sessions. If the data spread evenly across the 22 languages, each language would hold roughly 53 recordings and 4.9 hours. CIIL’s sources do not state the per-language split, so treat those two figures as back-of-envelope math, not as dataset facts.
The three recording conditions matter more than the headline hours. Diarization models often score well on clean studio audio and then collapse on phone-quality or noisy speech. A benchmark that deliberately mixes conditions pushes developers to report the hard case. The sources we reviewed do not name the three conditions, so we will not guess them.
A revision hash also signals good research hygiene. Benchmarks drift when maintainers fix labels or add files without telling anyone. Pinning 92877bad in a paper or a product evaluation lets a reader reproduce a result months later.
Licenses: Open Code, Mixed Terms on Weights #
Open-source speech AI is only as useful as its license allows. Anuvani’s code carries a dual license: MIT OR Apache-2.0. A user can pick either, and both permit commercial use with minimal conditions. The weights follow a different pattern, and that is where a product team must read carefully.
| Artifact | License listed | What it means in practice |
|---|---|---|
| Anuvani code (package, serving kit, reference implementation) | MIT OR Apache-2.0 | Permissive. Use, modify and ship under either license. |
| Segmentation weights | CC BY-NC 4.0 | Attribution required. Non-commercial use only unless CIIL grants more. |
| PLDA weights | CC BY 4.0 | Attribution required. Commercial use allowed. |
| Voiceprints | CC BY 4.0 | Attribution required. Commercial use allowed. |
| Indic DiarBench data | Not confirmed in sources reviewed | Check the dataset card before redistributing. |
The split creates a clear consequence. A startup can freely adopt the code and the PLDA and voiceprint components, but the segmentation network carries a non-commercial tag. A company that wants to run Anuvani-Diarization inside a paid product will need either permission from CIIL or a replacement segmentation model. We read the listed terms at face value, and a legal review should confirm that reading before anyone ships.
Upstream terms add one more layer. The segmentation and voiceprint networks come from the DiariZen and pyannote ecosystems, and those projects set their own conditions. The sources we reviewed do not spell those terms out, so teams should trace each checkpoint back to its origin and read its card.
What Is Confirmed and What Is Not #
Early coverage of any launch tends to blur reported facts with inference. The table below separates the two so readers can see where the evidence stops.
| Item | Status | Detail |
|---|---|---|
| CIIL launched Anuvani, an open-source speech-AI ecosystem | Confirmed | Reported by The Hindu and shown in the Hugging Face listing |
| Anuvani-LID and Anuvani-Diarization are open-weight models | Confirmed | Hugging Face listing |
| Coverage of 22 scheduled languages plus Indian English | Confirmed | The Hindu and Hugging Face |
| Indic DiarBench: 1,164 recordings, 107.7 hours, 22 languages, three conditions | Confirmed | Model card, revision 92877bad |
| Release 1.0.0 contents | Confirmed | anuvani_diar package, Triton and FastAPI kit, reference implementation |
| Exact launch date and venue | Unconfirmed | Not stated in sources reviewed |
| Development team or individual creators | Unconfirmed | Not stated in sources reviewed |
| Parameter counts and accuracy scores | Unconfirmed | Beyond dataset details, nothing verified |
| Pricing and commercial terms of the full ecosystem | Unconfirmed | Only the listed code and weight licenses are known |
No verified on-record quotation from a named CIIL speaker turned up in the sources available to us. That is why this article paraphrases the announcement and avoids attributing opinions to named officials. When CIIL or its researchers publish a paper or a detailed model card with metrics, the numbers in the unconfirmed rows will become checkable.
Why Indian Language Speech AI Has Been Hard #
India’s Constitution lists 22 scheduled languages, and the country’s linguistic reality runs well beyond that list. Speech technology has followed the money. Languages with the largest paying user bases got recognizers and voice assistants first, while smaller languages waited. This pattern repeats across the industry, and it is not unique to India.
Three structural problems compound the gap. Training data is scarce for many languages. Scripts and dialects vary widely inside a single language. And real speakers mix languages constantly. Product builders feel this daily. Our earlier report on India’s voice AI latency budget of 500 to 800 ms showed how little room a voice product has for slow or error-prone components when users expect a conversation-speed reply.
A public-sector language institute has a different set of incentives from a startup. CIIL does not need to hit a revenue target in the biggest language first. Its mandate points toward coverage, which is why the 22-language scope is the headline feature rather than a stretch goal. Whether the quality is even across all 22 languages is the question independent testers will answer.
Diarization has also lagged behind recognition in Indian languages, at least in public benchmarks. Transcription attracts more attention because users see the text. Diarization works in the background and gets noticed only when it fails, for example when a transcript attributes a customer’s complaint to the support agent. By publishing a dedicated benchmark, CIIL gives that quieter task a scoreboard.
How Anuvani Compares With Other Speech Models #
A fair comparison starts by admitting that these models do different jobs. Anuvani’s confirmed releases handle language identification and diarization. Nemotron 3.5 Transcribe handles transcription. They fit together in a pipeline more than they compete. The table lists what each component does and what the sources confirm about its scope.
| Component | Primary task | Language scope (confirmed) | Openness |
|---|---|---|---|
| Anuvani-LID | Spoken language identification | 22 scheduled languages plus Indian English | Open weights |
| Anuvani-Diarization | Who spoke when | 22 Indian languages in Indic DiarBench | Code MIT OR Apache-2.0, mixed weight licenses |
| Nvidia Nemotron 3.5 Transcribe | Speech transcription | 40 languages, per our earlier report | See our earlier report for terms |
| DiariZen WavLM-Large and Conformer | Segmentation inside Anuvani | Upstream component, not Indic-specific per the sources | Used under CC BY-NC 4.0 weights in Anuvani |
The open-weight angle links Anuvani to a wider shift. Our comparison of DeepSeek, Qwen and Mistral open-weight models tracked how quickly open releases now close gaps that once justified paying for closed APIs. Speech is following the same arc, and language-specific open models are where local institutions can add something global vendors have little reason to build.
One more difference separates Anuvani from most competitors. It ships a benchmark along with the model. Vendors with closed systems rarely publish evaluation sets that let outsiders test claims on a fixed revision. That transparency forces a more honest conversation about quality, which helps buyers even if they never run Anuvani in production.
Market Impact: Who Gains From Open Indian Speech Tooling #
The immediate winners are teams that build voice products for Indian users but lack the budget to collect speech data in 22 languages. Contact-center software vendors, meeting-note tools, media subtitling services, court and government transcription projects and education apps all need to know which language is spoken and who is speaking. Anuvani gives them a free starting point for both questions.
Commercial vendors face a more mixed picture. Companies that sell proprietary language identification or diarization services for Indian languages now compete against a public-sector alternative with open code. Some will respond by emphasizing accuracy guarantees, support contracts and integration work. Others will fold Anuvani into their own stacks where licenses allow it, which is a common pattern after strong open releases appear.
The non-commercial segmentation weights soften the immediate commercial impact. A paid product cannot simply drop them in. That leaves room for a licensing conversation with CIIL, for a vendor to train a commercial-friendly segmentation model, or for the institute to relax the terms in a later release. We cannot predict which path CIIL will choose, and the sources do not say.
Researchers gain the most clarity. A shared benchmark and an open reference implementation make it possible to publish comparable results on Indian-language diarization, and to build on each other’s work. For more coverage of this space, see our AI and machine learning section.
What Developers Can Do Today #
Pull the model and read the card
The fastest way to evaluate Anuvani-Diarization is to download the repository and read its model card before writing any code. The repository ID comes from the Hugging Face listing, and the standard command-line tool will fetch it.
pip install -U huggingface_hub
huggingface-cli download CIIL/Anuvani-Diarization --local-dir anuvani-diarization
Inside the download, look for the anuvani_diar package, the serving kit and the reference implementation. Run the reference code on a handful of your own recordings first. Then compare its speaker turns against a transcript you labeled by hand. Ten careful clips teach more than any leaderboard.
Check the license before you ship
Write down which weights your pipeline loads and map each one to the license in the table above. If a commercial product touches the segmentation weights, and contact CIIL or choose another segmentation model. Keep attribution text ready, because every listed weight license requires it.
For deployment, teams with GPU infrastructure can use the Triton path. Smaller teams can start with the FastAPI path on a single machine. If your stack leans on local inference, our guide to Ollama, LM Studio, llama.cpp and vLLM explains how those runtimes differ, though Anuvani’s own serving kit does not depend on any of them.
Risks and Limits Teams Should Weigh #
First, quality across languages may vary. A model that covers 22 languages can still perform unevenly, with strong results in languages that have more training audio and weaker results elsewhere. The sources we reviewed give no per-language scores, so no one outside CIIL can state how even the performance is.
Second, a benchmark of 107.7 hours is a useful start but a small sample for a country this diverse. Dialects, age groups, background noise and telephone codecs can all shift results. Teams should build a test set from their own audio and treat Indic DiarBench as one signal among several.
Third, voiceprints raise privacy questions. Any system that creates speaker embeddings processes biometric-like data. Under India’s data protection rules and any sector regulation that applies, a company should decide how long it stores embeddings, who can access them, and whether it needs consent. Diarization that only assigns anonymous labels within a single recording carries less risk than systems that match voices across recordings.
Fourth, the ecosystem is young. Release 1.0.0 means a stable first version, not a mature one. Teams should expect updates, pin versions in their builds and watch the repository for changes to weights or licenses.
The Sovereignty Angle: Who Builds India’s AI Foundations #
The launch lands in a live debate about national AI capacity. Elets CIO reported that Sarvam’s Vivek Raghavan warned India risks becoming a digital colony without its own AI foundations. Speech models for local languages sit near the bottom of that stack, because voice is how many Indians first meet digital services.
A national language institute releasing open speech tooling answers part of that worry. Foundations that the public can inspect and reuse reduce dependence on a few foreign vendors. They do not eliminate it. Training still needs compute, and the diarization pipeline rests on upstream networks from other research groups, as the model card shows.
Coverage questions reach beyond the biggest languages too. Sikkim Express ran a piece on AI, language technology and the Nepali language, a reminder that speakers of scheduled languages outside the major commercial markets watch these releases closely. Nepali is one of the 22 scheduled languages, and whether Anuvani serves its speakers well will show up in testing, not in a launch announcement.
Historical Context: From Closed Language Services to Open Models #
Speech technology for Indian languages passed through several phases. Early systems came from research labs and a few large companies and covered a handful of languages. Cloud APIs widened access but kept the models closed, so developers paid per minute and could not inspect failures. The recent wave of open-weight models changed the economics, because a team can now run, test and adapt a model on its own hardware.
Diarization followed a similar path in English first. Open toolkits such as the pyannote family made speaker labeling accessible to small teams, and the network named in Anuvani’s voiceprint stack comes from that world. What Indian languages lacked was an evaluation and adaptation layer built for their sounds and their recording conditions. Anuvani aims at that gap.
The institutional pattern also has precedent in other fields. Public bodies often fund the unglamorous infrastructure, such as corpora, benchmarks and reference tools, that private companies then use to build products. If Anuvani follows that pattern, its lasting value may come less from the first model weights and more from Indic DiarBench becoming the standard yardstick.
Predictions: Where Anuvani Goes Next #
These are our read of the situation, not reported facts. We label them as judgment calls.
- Independent evaluations will appear within months. Open weights and a pinned benchmark invite outside researchers to publish per-language results, and the weakest languages will draw the most scrutiny.
- The segmentation license will become the main commercial friction. Vendors will either negotiate with CIIL or train replacements. A relaxed license in a later release would speed adoption considerably.
- More Anuvani components will join the collection. The word ecosystem signals intent, and the natural neighbors of language identification and diarization are recognition and synthesis. We cannot say what CIIL will ship or when.
- Commercial voice platforms will wrap Anuvani into their pipelines. Teams that already fine-tune models, as in our walkthrough onfine-tuning an LLM with QLoRA , will likely adapt open speech components the same way.
- Indic DiarBench will become a reference point in papers. A fixed revision and 22-language coverage make it an easy citation, and citation habits tend to stick.
The takeaway: CIIL Anuvani gives Indian-language speech work a public, inspectable base layer. The strongest facts today are coverage of 22 scheduled languages plus Indian English, a 1,164-recording benchmark, and a release package developers can run. The biggest unknowns are accuracy, the team behind the work and the commercial terms.
FAQ #
What is CIIL Anuvani?
Anuvani is an open-source speech-AI ecosystem launched by the Central Institute of Indian Languages. Its first public models are Anuvani-LID and Anuvani-Diarization, and both are listed under the Anuvani Collection on Hugging Face.
Which languages does Anuvani support?
The models cover all 22 scheduled languages of India and Indian English. The Indic DiarBench evaluation data spans the same 22 Indian languages.
What does Anuvani-Diarization do?
It identifies who spoke when in multi-speaker audio. Developers use that output to attach speaker turns to transcripts or to separate speakers in call and meeting recordings.
Can companies use Anuvani in commercial products?
The code is licensed under MIT OR Apache-2.0, which permits commercial use. The segmentation weights use CC BY-NC 4.0, which restricts commercial use, while the PLDA and voiceprint weights use CC BY 4.0. The commercial terms of the broader ecosystem are not confirmed, so check each artifact and get legal advice before shipping.
What is Indic DiarBench?
It is a diarization dataset of 1,164 recordings totaling 107.7 hours across 22 Indian languages and three recording conditions. The listed dataset revision is 92877bad.
What does Release 1.0.0 include?
It includes the anuvani_diar Python package, a Triton Inference Server and FastAPI serving kit, and the original Python reference implementation.
How accurate are the Anuvani models?
The sources we reviewed do not provide verified accuracy figures, parameter counts or benchmark scores beyond the dataset details. Test the models on your own audio and watch for independent evaluations on Indic DiarBench.