IndicConformer

IndicConformer is AI4Bharat's open speech-recognition model for Indian languages, trained on real human speech datasets including Shruti and IndicSUPERB. Released under the permissive MIT license, it is one of the few openly available ASR checkpoints covering a wide range of Indic languages, and community conversions ship it in Sherpa-ONNX format for offline Android deployment at roughly 150MB.\n\nIt has become a practical baseline for Indic-language voice apps, appearing as the upstream checkpoint behind per-language fine-tunes and conversions for Hindi, Bengali, Marathi, Tamil, Kannada, Gujarati, Odia, and others.