Hugging Face · 2026-08-28 · notable
Open ASR Leaderboard adds Hindi — with Indian English sets from Voice Arena
The Open ASR Leaderboard added Hindi, its first Global South language, plus Indian English. Both evaluation sets come from a Voice Arena partnership and cover 4,888 speakers with 12 recorded speaker attributes per clip.
The main open speech-recognition leaderboard finally measures a language outside the usual Western set.
What is it?
Hindi is now scored on the Open ASR Leaderboard, the first Global South language it covers, alongside a new Indian English track. Both evaluation sets come from Voice Arena's Monsoon collection and are published on Hugging Face under CC-BY-4.0 as MonsoonASR-Open-ASR-leaderboard-hi-IN and -en-IN. Together the four splits hold 4,888 speakers of conversational, spontaneous speech.
How does it work?
Each clip carries 12 speaker attributes — gender, age, occupation, education, home state, device and more — so word error rate can be broken down by group rather than reported as one aggregate. Hindi references use a lattice of accepted orthographic variants instead of a single gold string, which stops a model being marked wrong for a legitimate spelling. The English references normalize disfluencies.
Why does it matter?
Aggregate scores hide who a speech model actually fails. In the launch analysis one model's word error rate moved 0.46 points across Indian regions while another moved 1.68 points, even though the two looked nearly identical on the overall number. Teams shipping voice products in India now have a public, license-clear way to see that gap.
Who is it for?
speech recognition researchers and voice product teams
Try it
https://huggingface.co/datasets/VoiceArena/MonsoonASR-Open-ASR-leaderboard-hi-IN