A non-profit based in India has quietly released a set of open-weight speech recognition models that could reshape how legal and government transcription gets done across South Asia. Adalat AI published SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR, a family of open models built for Hindi, Malayalam, and Kannada, according to reports citing the project’s research documentation. The release lands at a moment when most automatic speech recognition (ASR) progress is still concentrated in English and a handful of other high-resource languages, leaving a billion-plus speakers underserved by mainstream tools.
What makes SCRIBE notable isn’t just the language coverage. It’s the target use case: document-ready transcription, meaning output formatted with punctuation, capitalization, and structure clean enough to drop into a legal filing or government record without a human rewriting it from scratch. That’s a different bar than the word-level transcripts most ASR systems produce, and it’s the gap Adalat AI says its models are built to close.
What Adalat AI Actually Released
According to the SCRIBE research documentation, the release covers open-weight rich-transcription models for three languages: Hindi, Malayalam, and Kannada. The models were evaluated on two benchmarks built for this purpose: FLEURS-RO, used for general-purpose evaluation, and IN22-Legal, a domain-specific benchmark aimed at legal transcription accuracy. Pairing a general benchmark with a legal-domain one signals where Adalat AI expects these models to actually get used first: courtrooms, tribunals, and government offices processing spoken proceedings into written records.
The models are described in coverage of the release as being among the smallest open models built for rich-orthography transcription, with a confirmed size of 120 million parameters. For context, that’s small enough to run on modest hardware rather than requiring a data center GPU cluster, which matters enormously for government offices and courts in smaller Indian cities that don’t have access to high-end compute.
Behind the models sits a training method Adalat AI calls an LLM-based data-curation pipeline, paired with what the documentation describes as a reproducible recipe for building Indic rich-transcription systems. In plain terms: instead of relying purely on hand-labeled audio, the team used large language models to help clean, structure, and curate training data at scale, then published the recipe so other teams working on Indic languages (or other underserved language families) can replicate the approach.
SCRIBE Isn’t Adalat AI’s First Indic ASR Push
SCRIBE follows an earlier release from the same organization called Vividh-ASR, an open benchmark and open model set focused specifically on Hindi and Malayalam speech recognition. Vividh-ASR’s benchmark scale is reported at roughly 36 hours of Hindi audio and 26 hours of Malayalam audio, a modest but purpose-built dataset aimed at measuring real-world transcription accuracy rather than chasing leaderboard scale.
The two releases trace a pattern: Adalat AI is building out language coverage one benchmark and one model family at a time, rather than attempting a single sweeping multilingual release. Vividh-ASR’s blog post is dated August 28, 2026, while the SCRIBE arXiv version referenced in current reporting carries a June 30, 2026 date, putting roughly two months between the two publications. Read together, they show a non-profit iterating quickly on a narrow, well-defined problem instead of trying to out-scale labs with far bigger compute budgets.
Arghya Bhattacharya, identified in reporting as co-founder and CTO at Adalat AI, is named in connection with the organization’s work in this space. Beyond that attribution, specific on-record commentary about the SCRIBE release itself has not been published, so this piece sticks to what the documentation and reporting actually confirm rather than guessing at motive or roadmap.
Why Document-Ready Transcription Is the Hard Part
Most ASR systems, including well-known ones like OpenAI’s Whisper, produce a stream of words with limited structure. That’s fine for subtitles or search indexing. It’s not fine for a court record, where a transcript needs correct punctuation, speaker turns, capitalization of proper nouns, and formatting that holds up to legal scrutiny. Getting from “raw words” to “document-ready” usually requires a human editor, and that editing pass is often the single biggest bottleneck in digitizing court and government proceedings in India.
This is also where the IN22-Legal benchmark matters more than it might look at first glance. General ASR benchmarks measure word error rate against everyday speech. A legal-domain benchmark tests something narrower and arguably more commercially relevant: can the model hold up when a witness mispronounces a name, when a judge switches between Hindi and English mid-sentence (a routine occurrence in Indian courtrooms), or when technical legal terminology shows up that a general-purpose model has never seen in training.
Mainstream cloud ASR providers, including Google’s Speech-to-Text and Microsoft’s Azure AI Speech, do support Hindi and a growing list of Indian languages. What they don’t typically offer is an open-weight model tuned specifically for legal-document output that a court system can self-host, audit, and modify without sending sensitive proceedings data to a third-party cloud API. That self-hosting angle is likely to be the real selling point for government adoption, more than raw accuracy numbers.
The Open-Weight Angle: Why It Matters for Government Adoption
Releasing SCRIBE as open weights rather than as a hosted API is a deliberate choice with real consequences. Government procurement, especially around justice-system infrastructure, tends to move slowly and carries hard requirements around data sovereignty. A court system that cannot legally send recordings of in-camera proceedings to a foreign cloud provider has very few options if the only available ASR tools are closed, hosted services.
By publishing open weights alongside a documented training recipe, Adalat AI is lowering the barrier for any state judiciary, research lab, or civic-tech group to download the models, run them on local infrastructure, and adapt them to a specific court’s needs, whether that’s a different regional dialect, a specialized vocabulary, or integration into an existing case-management system. This mirrors a broader trend across AI releases in 2026, where open-weight models from across the industry are increasingly positioned as the practical choice for regulated, data-sensitive deployments rather than as a budget alternative to closed models.
That trend shows up elsewhere in the AI landscape too. DeepSeek’s V4.1-Flash release cut output pricing by 70%, and MiniCPM5-2B demonstrated that small, efficient open models can beat larger proprietary ones on specific benchmarks. Adalat AI’s 120M-parameter SCRIBE models fit the same pattern: small, efficient, purpose-built, and open, rather than enormous and general.
Competitive Landscape: How SCRIBE Stacks Up
It’s worth being precise about what can and can’t be compared here. Adalat AI has not published pricing, and no independent head-to-head benchmark against Whisper, Google Speech-to-Text, or Azure AI Speech on Hindi, Malayalam, or Kannada has been confirmed in available reporting. What can be compared is scope, licensing model, and target use case.
| System | Model type | Target languages (Indic) | Self-hostable | Legal-domain tuning |
|---|---|---|---|---|
| Adalat AI SCRIBE | Open-weight, 120M params | Hindi, Malayalam, Kannada | Yes | Yes (IN22-Legal benchmark) |
| Adalat AI Vividh-ASR | Open benchmark + model | Hindi, Malayalam | Yes | No (general benchmark) |
| OpenAI Whisper | Open-weight, multiple sizes | Broad multilingual, incl. Hindi | Yes | No |
| Google Speech-to-Text | Hosted API | Multiple Indian languages | No | No |
| Azure AI Speech | Hosted API | Multiple Indian languages | No | No |
The pattern is clear: the big cloud providers cover more languages and more infrastructure scale, but none of them offer an open, self-hostable, legal-domain-tuned model for Indic languages. That’s the specific niche SCRIBE occupies, and it’s a niche none of the larger players have prioritized, presumably because the commercial market for Indic legal transcription is small relative to the engineering effort required to serve it well.
Benchmark Design: FLEURS-RO vs IN22-Legal
The choice to evaluate SCRIBE on two separate benchmarks rather than one tells its own story about what Adalat AI is optimizing for. FLEURS-RO functions as the general-purpose stress test, likely built on the broader multilingual FLEURS evaluation framework that’s become a common reference point across the ASR field, extended here with rich orthography requirements like punctuation and capitalization rather than raw word sequences. IN22-Legal narrows the focus to the exact register SCRIBE is meant to serve: formal, structured legal speech where a missed comma or an uncapitalized name can change how a transcript reads in a filed document.
| Benchmark | Purpose | Speech register | Primary metric focus |
|---|---|---|---|
| FLEURS-RO | General ASR evaluation | Everyday/broad speech | Rich-orthography accuracy |
| IN22-Legal | Domain-specific evaluation | Formal legal proceedings | Document-ready transcription accuracy |
| Vividh-ASR benchmark (Hindi) | Prior general benchmark | Everyday speech | ~36 hours of labeled audio |
| Vividh-ASR benchmark (Malayalam) | Prior general benchmark | Everyday speech | ~26 hours of labeled audio |
Running both a general and a domain-specific benchmark side by side also gives outside researchers a way to sanity-check whether SCRIBE’s legal-domain tuning actually trades off against general accuracy, a common failure mode in domain-specialized models. Publishing both scores rather than just the flattering one is itself a small but meaningful transparency choice, and it puts SCRIBE in different territory than most commercial ASR marketing, which tends to cite a single favorable number without the comparison context.
How This Fits Into the Broader Open-Model Moment
SCRIBE’s release doesn’t happen in a vacuum. 2026 has seen a steady stream of open and open-weight model releases positioned around specific, practical niches rather than general-purpose dominance. Mistral opened up an AI safety tool alongside its valuation news, Fastino shipped a 340-million-parameter model tuned to run efficiently on CPU rather than requiring a GPU, and GLM-5.3’s open weights have been scrutinized heavily for both capability and risk. Against that backdrop, a 120-million-parameter model built for a narrow, well-defined task like legal transcription looks less like an outlier and more like where a meaningful slice of the open-model ecosystem is actually heading: smaller, cheaper to run, and purpose-built rather than general.
That’s a different trajectory from the frontier-model race playing out among the largest labs, where releases like Gemini 4 Argon and Claude Opus 5.5 compete on context length, raw benchmark scores, and enterprise pricing. Adalat AI isn’t trying to compete on that axis at all. SCRIBE’s value proposition is narrower and arguably more durable: a specific, underserved problem, solved well enough to be usable, and released in a form that the people who need it most (cash-strapped public institutions) can actually deploy without a six-figure cloud contract.
Historical Context: India’s Long Road to Digitized Court Records
India’s judiciary has been pursuing digitization for over a decade through initiatives like the e-Courts project, which aimed to computerize case filings, hearing schedules, and judgments across the country’s district courts. Transcription of actual spoken proceedings, though, has lagged far behind document digitization, largely because of the sheer linguistic diversity involved. India recognizes 22 scheduled languages, and court proceedings routinely mix regional languages with Hindi and English within a single hearing.
Academic and civic-tech efforts to build Indic speech datasets go back years, with projects like Mozilla’s Common Voice crowdsourcing voice data across dozens of languages, including several Indian ones. What’s been missing until recently is a model specifically tuned for the structured, formal register of legal speech rather than everyday conversation. SCRIBE and Vividh-ASR are part of a wave of 2026 releases attempting to close that specific gap, building on years of underlying dataset work rather than starting from zero.
Market Impact: A Narrow but Real Opportunity
The market for legal and government transcription tools in India is not the kind of headline-grabbing, billion-dollar opportunity that dominates AI coverage elsewhere. It’s a narrower, public-sector-driven market, shaped by procurement cycles, data-residency rules, and budget constraints at the state level rather than venture funding rounds. But narrow doesn’t mean small in absolute terms: India’s district and subordinate courts handle tens of millions of pending cases, and every one of those cases generates spoken proceedings that currently rely heavily on manual stenography or post-hearing summarization by clerks.
If even a fraction of that backlog shifts toward automated, document-ready transcription, the efficiency gain for court administration could be substantial, even without any single large commercial contract driving it. The open-weight nature of SCRIBE also means Adalat AI isn’t necessarily positioning itself to capture that value directly through licensing fees. Instead, the more likely path to sustainability for a non-profit releasing open models is service contracts, customization work, or grant funding tied to public-interest technology, a model common among civic-tech organizations working in the legal-tech space globally.
Technical Approach: The LLM-Based Curation Pipeline
One of the more transferable contributions from SCRIBE isn’t the models themselves but the data pipeline behind them. Using large language models to curate and structure training data, rather than relying solely on manual annotation, has become increasingly common across the AI field in 2026 as teams look for ways to stretch limited labeled-data budgets further. For low-resource languages, where manually annotated audio-text pairs are scarce and expensive to produce, this kind of LLM-assisted curation can meaningfully lower the cost of building a usable training set.
Because Adalat AI published this as a reproducible recipe rather than keeping it proprietary, teams working on other underserved languages, Indic or otherwise, have a documented starting point rather than having to reinvent the approach from scratch. That’s arguably a bigger long-term contribution than the three specific language models released this time, since recipes compound across projects in a way that any single model release doesn’t.
What’s Missing From the Public Record
To be transparent about the limits of current reporting: Adalat AI’s pricing (if any licensing or service fees apply to deployment support), direct on-record commentary from the team about SCRIBE specifically, and the organization’s formal legal or non-profit registration status have not been confirmed in available documentation. Readers should treat those details as open questions rather than settled facts until Adalat AI or independent outlets publish more specifics. This article sticks to what’s documented: the model architecture, the benchmark design, the language coverage, and the prior Vividh-ASR release that preceded it.
Predictions: Where Indic ASR Goes From Here
- Expect SCRIBE’s language coverage to expand beyond Hindi, Malayalam, and Kannada within the next year, following the same incremental pattern set by the Vividh-ASR to SCRIBE progression.
- Other civic-tech and academic groups working on underserved languages will likely adopt or adapt Adalat AI’s LLM-based curation pipeline, given how expensive manual annotation remains for low-resource languages.
- State-level judiciary bodies in India are the most plausible first adopters, given the direct fit between SCRIBE’s legal-domain tuning and existing e-Courts digitization efforts.
- Pressure will grow on major cloud ASR providers to either improve Indic-language accuracy or offer more flexible self-hosting options, as open alternatives close the usability gap for public-sector buyers.
- Expect more benchmark-first releases like FLEURS-RO and IN22-Legal across other regional language families, since publishing a rigorous evaluation set alongside a model has become a credibility signal in the open-model ecosystem.
Frequently Asked Questions
What is Adalat AI’s SCRIBE model?
SCRIBE is a set of open-weight speech recognition models built by Adalat AI for rich, document-ready transcription in Hindi, Malayalam, and Kannada, evaluated on the FLEURS-RO and IN22-Legal benchmarks.
How big are the SCRIBE models?
Reporting on the release confirms a 120-million-parameter size, described as among the smallest open models built for rich-orthography transcription.
What is IN22-Legal?
IN22-Legal is a domain-specific benchmark used to evaluate SCRIBE’s transcription accuracy on legal-proceeding speech, separate from the general-purpose FLEURS-RO benchmark.
Is SCRIBE Adalat AI’s first ASR release?
No. Adalat AI previously released Vividh-ASR, an open benchmark and model covering Hindi and Malayalam, with a blog post dated August 28, 2026, ahead of the SCRIBE arXiv version dated June 30, 2026.
Can SCRIBE be self-hosted?
Yes. Because the models are released as open weights with a documented training recipe, organizations can download and run them on their own infrastructure rather than depending on a hosted API.
How does SCRIBE compare to Whisper or Google Speech-to-Text?
Whisper and the major cloud ASR APIs support broader multilingual coverage but are not specifically tuned for document-ready legal transcription in Indic languages, and the cloud APIs are not self-hostable, which matters for data-sensitive government use cases.
Who is behind Adalat AI?
Arghya Bhattacharya is identified in reporting as co-founder and CTO at Adalat AI. Further organizational details, including its formal non-profit status, have not been independently confirmed.
Is SCRIBE free to use?
Pricing details have not been published. As an open-weight release, the models themselves can be downloaded without a licensing fee, though no information is currently available on paid support or deployment services.
