
🚨Indic ASR Drop from @Adalat_AI 🚨 We built a 244M ASR model that beats models up to 6x its size on Hindi and Malayalam. The secret? We broke the standard fine-tuning playbook. 1. Instead of easy → hard, we did hard → easy. 2. High LR to shatter the decoder’s pretrained basin. 3. Reverse curriculum so peak plasticity hits the hardest acoustic conditions. The results :- → Hindi: 15.73 WER vs 21.05 (Vaani Large-v3, 1.5B by @artparkindia ) → Malayalam: 39.64 vs 47.96 (IndicWhisper, 769M by @ai4bharat) Releasing Vividh-ASR + model weights openly. Link in comments. Credits to @kavya_manohar, @kushjuvekar & Manas for shipping this in service of the mission!





















