{
  "strict_load": true,
  "source_revision": "2d6c1f99c129b3ef07a8d168bc0ff08e2bdb93f1",
  "device": "cpu",
  "dtype": "torch.bfloat16",
  "feature_extractor": "official 16 kHz log-mel extractor",
  "tests": [
    {
      "input": "synthetic 440 Hz sine, 0.5 s",
      "samples": 8000,
      "feature_shape": [
        1,
        128,
        49
      ],
      "encoder_shape": [
        1,
        7,
        768
      ],
      "encoder_lengths": [
        7
      ],
      "projector_shape": [
        7,
        2048
      ],
      "finite": true,
      "repeat_max_abs_error": 0.0
    },
    {
      "input": "public GRT-Omni generated speech sample, 0.95 s",
      "samples": 15200,
      "feature_shape": [
        1,
        128,
        94
      ],
      "encoder_shape": [
        1,
        12,
        768
      ],
      "encoder_lengths": [
        12
      ],
      "projector_shape": [
        12,
        2048
      ],
      "finite": true,
      "repeat_max_abs_error": 0.0
    }
  ],
  "speech_sample_source": "https://huggingface.co/MatterWaveAI/grt-30m-omni/blob/c69d1331739458fd92b1d3038ed685a13840a545/sample_omni_speech.wav",
  "finite": true,
  "elapsed_seconds": 4.16,
  "limitations": "Component-only CPU smoke tests; speech sample is generated, not a human recording. No recognition quality, full-model parity or GPU test."
}