60dB delivers the lowest
Word Error Rate on Hindi

• Datasets: Hindi eval splits of SkunkWorkLabs/hindi-asr-benchmark and RinggAI/ASR-Benchmarking-Dataset — six subsets (Common Voice, FLEURS, IndicTTS, Kathbath, Kathbath-noisy, MUCS).
• Sample: every record in each dataset — 9,997 clips total, no sampling.
• Metric: Word Error Rate (WER), lower is better, aggregated by total reference words (micro-average) so longer clips contribute proportionally.
• 60dB: transcribed live through our production APIs — both the HTTP/batch endpoint and the WebSocket/streaming endpoint — using the language hint hi,en.
• Other providers: word error rates are taken from the datasets' own pre-computed, normalized vendor columns (we did not re-run those services); they reflect each vendor's result at the time the dataset authors ran them.
• Coverage: 60dB (HTTP) 9,997/9,997; vendor columns 9,914–9,996/9,997 (a handful of blank cells per provider).

Notes & disclosures

1. 1Vendor numbers are dataset-provided, not produced by us, and may reflect older API versions of those services.
2. 2The RinggAI dataset is published by Ringg AI — a competitor's own benchmark — which can favour their reference conventions. 60dB still leads it overall.
3. 3FLEURS vendor columns are partly corrupted; we report it transparently but exclude it from the headline ranking.
4. 4Common Voice (clean read-speech) is our weakest subset. Raw WER there also penalises rendering convention — 60dB writes common English loanwords in Latin script (e.g. branch manager, image) where the reference uses Devanagari; much of that gap is stylistic, not accuracy.
5. 5Streaming (WebSocket) vs batch (HTTP): the real-time streaming path scores slightly higher WER than batch (expected — it commits words incrementally under latency constraints) yet still ranks at or near the top among all providers.