{
  "model": {
    "name": "stt_en_conformer_ctc_small",
    "source": "NVIDIA NeMo EncDecCTCModelBPE",
    "onnx": "exports/models/stt_en_conformer_ctc_small_onnx/stt_en_conformer_ctc_small_features.onnx",
    "onnxBytes": 53197388,
    "opset": 17,
    "dummyFeatureFrames": 6000
  },
  "audio": {
    "sampleRate": 16000,
    "channels": 1,
    "sampleFormat": "float32 PCM in [-1, 1]"
  },
  "features": {
    "type": "log-mel-spectrogram",
    "features": 80,
    "normalize": "per_feature",
    "windowSizeSeconds": 0.025,
    "windowStrideSeconds": 0.01,
    "window": "hann",
    "nFft": 512,
    "log": true,
    "frameSplicing": 1,
    "dither": 1e-05,
    "padTo": 0,
    "padValue": 0.0
  },
  "ctc": {
    "tokenizer": "SentencePiece",
    "tokenCountWithoutBlank": 1024,
    "tokenCountWithBlank": 1025,
    "blankId": 1024,
    "featureFrameStrideSeconds": 0.01,
    "encoderSubsampling": 4,
    "approxLogitFrameStrideSeconds": 0.04
  },
  "onnxContract": {
    "inputs": [
      {
        "name": "processed_signal",
        "shape": [
          "batch",
          80,
          "feature_frames"
        ],
        "dtype": "float32"
      },
      {
        "name": "processed_signal_length",
        "shape": [
          "batch"
        ],
        "dtype": "int64"
      }
    ],
    "outputs": [
      {
        "name": "log_probs",
        "shape": [
          "batch",
          "encoded_frames",
          1025
        ],
        "dtype": "float32"
      },
      {
        "name": "encoded_len",
        "shape": [
          "batch"
        ],
        "dtype": "int64"
      }
    ]
  },
  "tokenizerAssets": {
    "vocab.txt": "exports/models/stt_en_conformer_ctc_small_onnx/tokenizer/vocab.txt",
    "tokenizer.model": "exports/models/stt_en_conformer_ctc_small_onnx/tokenizer/tokenizer.model",
    "tokenizer.vocab": "exports/models/stt_en_conformer_ctc_small_onnx/tokenizer/tokenizer.vocab",
    "model_config.yaml": "exports/models/stt_en_conformer_ctc_small_onnx/tokenizer/model_config.yaml"
  },
  "mobileRuntimeNotes": [
    "Compute features with the exact parameters above before calling ONNX.",
    "Use the exported SentencePiece tokenizer; do not build a custom a-z tokenizer.",
    "Run constrained CTC Viterbi in native code; do not use beam search for known transcript alignment."
  ]
}