• Eight models land between 4.81 and 4.99 WER on the Indian English public set, a 0.18-point band, so aggregate WER does not separate them.
• Regional spread does separate them: Whisper-large-v3-turbo varies by 0.46 points across five zones, Voxtral-Mini-3B-2507 by 1.68 (4.38 WER in the Central zone, 6.06 in the East).
• Hindi is scored with OIWER, which accepts any linguist-validated spelling variant instead of charging it as an error.
• The Hindi public split is only 1.33 hours, so self-scoring there rests on a thin sample; the 4.47-hour Hindi private split is withheld.
• The post text we reviewed gives no full per-model table, so no Hindi ranking can be stated from it.

The Open ASR Leaderboard now scores Hindi and Indian English, and the useful signal is not who ranks first. In a post dated August 28, 2026, Hugging Face and Voice Arena describe eight models landing between 4.81 and 4.99 WER on the Indian English public split, a 0.18-point band. What separates the models is how evenly they perform across regions: Whisper-large-v3-turbo varies by 0.46 points across five zones, Voxtral-Mini-3B-2507 by 1.68. The bottom line is Use it, with the spread as the deciding number.

Use it: the badge and the number behind it

The Weights' call is Use it, because the sets add a way to see regional spread and orthographic variance in Indian speech. The deciding number is 1.68 versus 0.46. On the Indian English public set, Voxtral-Mini-3B-2507 scores 4.38 WER in the Central zone and 6.06 in the East, while Whisper-large-v3-turbo moves 0.46 points across the same five zones, according to the post. The badge attaches to the benchmark, not to any model, and it is conditional on reading spread instead of rank.

Four speaker-disjoint splits, with a thin Hindi public set

The release comprises four speaker-disjoint splits, a public and a private one for each language. Indian English has 5.62 hours from 1,444 speakers public and 5.58 hours from 1,405 speakers private. Hindi has 1.33 hours from 468 speakers public and 4.47 hours from 1,571 speakers private. The public splits are for self-scoring; the private ones are withheld to limit benchmark-specific optimisation, per the post. At 1.33 hours, the Hindi public split is the smallest, so a team scoring itself there is working from a thin sample.

Eight models inside 0.18 WER points cannot be ranked by score

A 0.18-point band is too narrow to choose on. The post says eight models on the leaderboard land between 4.81 and 4.99 WER on the Indian English public set. The text we reviewed neither names all eight nor gives their exact scores, and we found no confidence intervals, so we cannot say which models sit at each end or whether the gaps are meaningful. What the band does suggest is that headline WER is close to indistinguishable here, and a deployment decision has to turn on something else: latency budget, cost per call, or consistency across regions.

Regional spread is the failure mode an average hides

The zone breakdown is where the models separate. Whisper-large-v3-turbo varies by 0.46 points across five geographic zones on the Indian English public set. Voxtral-Mini-3B-2507 varies by 1.68, from 4.38 WER in the Central zone to 6.06 in the East. That is roughly 38 percent more word errors in the East than in the Central zone for one model, and the spread is more than nine times the 0.18-point band that separates all eight models. A single averaged score would bury that gap.

OIWER stops charging models for valid Hindi spelling

Hindi is scored with OIWER, Orthographically-Informed Word Error Rate, rather than plain WER. Per the post, the metric scores a hypothesis against the accepted set of written forms at each span of the transcript, so any admitted form counts as correct and only genuine recognition errors are charged. Candidate variants are drawn from multiple ASR transcripts of the same audio, expanded with language models, and then native-speaker linguists decide which are valid for that utterance. The post cites a paper for the method and releases scoring code as the voi-oiwer package.

What the sets do not fix

The authors say the sets do not fix orthographic and demographic evaluation problems; what they add is a way to see them. Two limits follow. The post says the Hindi sets concentrate regionally more tightly than the Indian English ones, so zone spread measured on English cannot be assumed to transfer to Hindi. And the text we reviewed reports no per-model Hindi OIWER table, so no Hindi ranking can be stated here. Teams choosing on Hindi should score their own audio.

The bottom line

The bottom line: Use it as a robustness check, not a shopping list. Test a shortlist for regional consistency rather than copying someone else's model choice. Score candidates per zone on Indian English and with OIWER on Hindi, and treat a spread near Voxtral's 1.68 points as a red flag if your users span regions. Then confirm on your own audio, since the private splits are withheld and the Hindi public split is only 1.33 hours. Watch for per-zone tables and a full Hindi OIWER ranking on the leaderboard.