VoiceArena/MonsoonASR-Open-ASR-leaderboard-hi-IN

Voice Arena Monsoon hi (public test)

Part of the Open ASR Leaderboard, on the Multilingual tab, where a model is ranked only if it supports every selected language.

A conversational Hindi ASR test set that records who was speaking, not only what was said. Every clip carries twelve speaker attributes — gender, age, native district and state, education, occupation, income band, handset — so a difference between two systems can be traced to a group of speakers instead of disappearing into one number.

This is the public split, released for self-scoring. A speaker-disjoint private split is withheld to limit benchmark-specific optimisation.

At a glance

Clips	753
Duration	1.33 h
Speakers	468
Clip length	6.4 s mean / 5.0 s median
Male / female clips	54 / 46
Native districts	202
States / UTs	11 states + 3 union territories
Distinct handset models	315
Current cities	289

Small in hours, large in speakers: more than half the speakers contribute exactly one clip, and the ten largest contributors account for under 7 per cent of total duration. A result here is an average over hundreds of voices rather than a few talkers recorded at length.

Loading

from datasets import load_dataset

ds = load_dataset("VoiceArena/MonsoonASR-Open-ASR-leaderboard-hi-IN", split="test")
print(ds[0]["lattice"], ds[0]["native_district"], ds[0]["gender"])

Audio is 16 kHz.

Scoring

lattice is a list of lists: one list of accepted written forms per span of the transcript. Hindi has no single correct orthography — everyday speech is code-mixed, English-origin words have no settled Devanagari spelling, and compounds are written joined or separated by preference — so a single-string reference would score systems on the annotator's spelling choices as much as on recognition.

Score with OIWER, which aligns a hypothesis against the accepted set at each span so any admitted form counts as correct:

pip install voi-oiwer

from voi_oiwer import oiwer

result = oiwer(hypothesis, example["lattice"], "hindi")
score, (ins, dele, sub), ref_words = result[0], result[4], result[5]

Report the aggregate over the split — sum(ins + del + sub) / sum(ref_words) — not the average of per-clip rates. Implementation: voi-oiwer.

Fields

Group	Fields
Segment	id, audio, audio_length_s, language
Reference	lattice
Speaker	speaker_id, gender, date_of_birth
Background	occupation, educational_background, marital_status, income
Geography	native_district, native_state, current_city, years_spent_in_current_district
Recording	device_manufacturer, device_model

More

Collection, quality control, the transcription protocol and a worked regional analysis: https://huggingface.co/blog/open-asr-leaderboard-global-south.