Medocr Asr Bn

--- license: mit language: - bn tags: - automatic-speech-recognition - onnx - bangla --- # medocr-asr-bn — on-device Bangla ASR (ONNX) for the MedOCR pharmacy app INT8 (MatMul-only dynamic quantization) ONNX export of the encoder of [ai4bharat/indicconformer_stt_bn_hybrid_ctc_rnnt_large](https://huggingface.co/ai4bharat/indicconformer_stt_bn_hybrid_ctc_rnnt_large) (MIT license, AI4Bharat), plus its CTC head, vocabulary, and mel filterbank. MatMul-only quantization is used because onnxruntime-android has no ConvInteger kernel. Used by the MedOCR Android app for offline voice stocking (the pharmacist speaks price / expiry / quantity). The app downloads `asr_encoder_bn.onnx` once on first use; the small files ship inside the APK. | file | bytes | |---|---| | asr_encoder_bn.onnx | 185341627 | | asr_ctc_bn.onnx | 11559655 | | asr_vocab_bn.json | 57466 | | asr_mel_filters.bin | 82240 | Contract: encoder `audio_signal [1,80,T] float32` + `length [1] int64` → `[1,512,T']`; ctc → logprobs `[1,T',5633]`; aggregate SentencePiece vocab, blank id 5632; 16 kHz mono input.

Source context: 0 downloads · 0 likes · Pipeline automatic-speech-recognition · Repo MrJAJA/medocr-asr-bn