{
  "model": "silero-vad-v5",
  "format": "tflite",
  "sample_rate": 16000,
  "context_samples": 64,
  "chunk_samples": 512,
  "total_samples": 576,
  "state_shape": [
    2,
    1,
    128
  ],
  "inputs": {
    "audio": {
      "shape": [
        1,
        576
      ],
      "dtype": "float32"
    },
    "state": {
      "shape": [
        2,
        1,
        128
      ],
      "dtype": "float32"
    }
  },
  "outputs": {
    "probability": {
      "shape": [
        1,
        1
      ],
      "dtype": "float32"
    },
    "state_out": {
      "shape": [
        2,
        1,
        128
      ],
      "dtype": "float32"
    }
  },
  "notes": "Callers must prepend the last 64 samples of the previous chunk to each new 512-sample chunk before inference. On the first chunk, pass zeros for the context."
}