{
  "display_name": "bengali-asr",
  "model_id": "bengali-asr-hybrid-1.0",
  "version": "1.0.0",
  "original_source": "ai4bharat/indicconformer_stt_bn_hybrid_ctc_rnnt_large",
  "original_model": "indicconformer_stt_bn_hybrid_ctc_rnnt_large",
  "description": "bengali-asr 1:1 verbatim ONNX — same Conformer-L 120M 17x512 5633 vocab (22*256+blank) + preprocessor wav->logits, rebranded folder but weights/config verbatim for audit. Use with language_id='bn' (Bengali) or any of 22.",
  "language_primary": "bn",
  "languages_all": [
    "as",
    "bn",
    "brx",
    "doi",
    "kok",
    "gu",
    "hi",
    "kn",
    "ks",
    "mai",
    "ml",
    "mr",
    "mni",
    "ne",
    "or",
    "pa",
    "sa",
    "sat",
    "sd",
    "ta",
    "te",
    "ur"
  ],
  "vocab_per_lang": 256,
  "vocab_total": 5632,
  "vocab_plus_blank": 5633,
  "blank_id": 5632,
  "blank_is_shared": true,
  "sample_rate": 16000,
  "architecture": "Conformer-L 120M (17 blocks, 512 dim, 8 heads, subsample 4) + RNNT decoder 640 + joint 640 + aux CTC 5633",
  "preprocessor": "AudioToMelSpectrogramPreprocessor n_fft 512 win 400 hop 160 mel 80 log+per_feature (inside ONNX)",
  "decoders": [
    "ctc",
    "rnnt"
  ],
  "input_wav": "wav [B,T] float32 16k mono + wav_len [B] int64 samples",
  "output_logits": "ctc_logits [B,T/4,5633] float32 (use language mask for bn: 256+blank)",
  "providers": [
    "CUDAExecutionProvider",
    "CPUExecutionProvider",
    "OpenVINOExecutionProvider",
    "CoreMLExecutionProvider"
  ],
  "export_date": "2026-08-26",
  "onnx_opset": 18,
  "rebrand_note": "Folder/models renamed to bengali-asr for new identity, but ONNX metadata keeps original_source for audit; weights are verbatim copy (no slicing like BanglaNeo 257)."
}