Redistribution-for-mobile bundle of AI4Bharat IndicConformer-600M (ai4bharat/indic-conformer-600m-multilingual, MIT), trimmed to the Hindi RNNT decode path with the encoder dynamically quantized to int8 for on-device iOS use (~698 MB; onnxruntime, CPU/CoreML).

Files (see manifest.json for sizes + sha256): encoder.onnx (int8), rnnt_decoder.onnx, joint_enc.onnx, joint_pred.onnx, joint_pre_net.onnx, joint_post_net_hi.onnx, vocab.json. Frame step 0.08 s; greedy RNNT decode emits word timings from the encoder-frame index. Original model © AI4Bharat, MIT. Quantization/repackaging only.