Quantized (int8, dynamic) ONNX Runtime export of ARTPARK-IISc/SraVaani-1.0, a ~430M parameter FastConformer ASR model with a hybrid TDT-CTC decoder covering 65 Indian languages and dialects. Produced for on-device (mobile) inference, where the original fp32 ONNX export (~1.7GB encoder alone) is too large to ship.

All credit for the underlying model, training, and the original ONNX export goes to ARTPARK-IISc — see their model card for training data, supported languages, and evaluation results. This repo only adds int8 dynamic quantization (via onnxruntime.quantization.quantize_dynamic) on top of their published ONNX files, redistributed under the same MIT license.

MIT, same as the original model.