Sarvam AI released Saaras V4, a speech recognition model covering all 22 scheduled Indian languages plus English, available through its API at ₹30 per hour. The encoder-decoder system uses Sarvam-3B, a 3B-parameter hybrid state-space decoder trained in-house. Sarvam reports state-of-the-art accuracy across the 22 languages, and 16.03% WER on IndicContextEval L5 with keyterm prompting. All benchmarks are vendor-reported; independent reproduction has not been published. Weights are not public—self-hosting docs cover Saaras v3 only.
No score is assigned. Sources and their independence are shown in the citation chain below.