syncvsr_export_metadata.json

{
  "model_name": "syncvsr_lrs3_visual_ctc",
  "source_checkpoint": "Vox+LRS2+LRS3.ckpt",
  "input_layout": "NTCHW",
  "input_channels": 1,
  "input_height": 88,
  "input_width": 88,
  "pixel_mean": 0.421,
  "pixel_std": 0.165,
  "output_layout": "NTV",
  "output_is_log_softmax": true,
  "blank_index": 0,
  "vocab_file": "syncvsr_unigram_units.txt",
  "notes": "Exported by tools/export_syncvsr_to_onnx.ipynb. Encoder + CTC head only -- no attention decoder, no beam search, no LM. Decode greedily or with subword CTC beam search; KenLM rescoring on top via KenLmScorer."
}