Datasetpublic

adalat-ai/in22-legal

IN22-Legal Test-only out-of-distribution legal-domain dictation benchmark for Indic ASR. Read-speech recordings of legal passages from the IN22-Gen corpus, dense in domain entities (statute names, section numbers), formal numerals (dates, monetary amounts), and complex clause structures. Released alongside the SCRIBE evaluation framework in SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR (Interspeech 2026, under review). 📄 Documentation: DATASHEET.md… See the full description on the dataset page: https://huggingface.co/datasets/adalat-ai/in22-legal.

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face

0likes42downloads

Dataset Card

IN22-Legal

Test-only out-of-distribution legal-domain dictation benchmark for Indic ASR. Read-speech recordings of legal passages from the IN22-Gen corpus, dense in domain entities (statute names, section numbers), formal numerals (dates, monetary amounts), and complex clause structures.

Released alongside the SCRIBE evaluation framework in SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR (Interspeech 2026, under review).

📄 Documentation: `DATASHEET.md` (Gebru-style datasheet documenting motivation, composition, collection, and maintenance).

Languages and splits

Language	Config	Examples	Total audio	Mean clip	Speakers
Malayalam	`ml`	80	21.4 min	16.1 s	4
Hindi	`hi`	80	33.7 min	25.3 s	2
Kannada	`kn`	80	20.1 min	15.1 s	2

Only a test split is provided.

Schema

Column	Type	Description
`audio`	Audio	48 kHz mono
`transcription`	string	Rich-transcription reference (legal-grade target)
`speaker_id`	string	Per-recording speaker identifier
`duration`	float	Audio duration in seconds
`language`	string	`Malayalam` / `Hindi` / `Kannada`

Curation

Legal passages were drawn from the IN22 corpus and recorded as read speech by 2–4 native speakers per language (Malayalam: 4, Hindi & Kannada: 2). Ground-truth transcripts were prepared directly in rich-transcription format — grammatical punctuation, formatted numerals (dates, statute/section references, monetary amounts), and Indic-script orthographic conventions — and manually verified by legal-domain annotators on a per-sample basis. The transcripts in this release are human-corrected, not LLM outputs.

Intended use

Out-of-distribution evaluation for ASR systems trained on general-domain corpora. Because none of the SCRIBE training data contains legal text, IN22-Legal tests whether formatting conventions learned from general corpora generalize to high-stakes specialized vocabulary.

Usage

python

from datasets import load_dataset

# pass "ml" / "hi" / "kn" as the config
ds = load_dataset("adalat-ai/in22-legal", "ml", split="test")
print(ds[0]["transcription"])

Evaluation

Use SCRIBE for diagnostic categorical evaluation (WER / LER / NER / PER / TER). The LER (Legal Error Rate) category surfaces failures on legal entities — statute names, section references, witness designations — that are otherwise drowned out in monolithic WER.

Sister benchmark for general-domain rich transcription: `adalat-ai/fleurs-ro`.

Limitations

Each language has 80 read-speech samples from 2–4 speakers, so speaker- and topic-coverage are limited by design. The benchmark is intended as a diagnostic out-of-distribution probe, not a saturation-level test of legal ASR.

Citation

—Source text passages: Gala et al., *IndicTrans2 / IN22-Gen*, 2023 (`ai4bharat/IN22-Gen`).
—Recording, rich-transcription annotation, and benchmark release: SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR, Interspeech 2026 (under review).

License

CC-BY-4.0.