{
  "schema_version": 1,
  "source": {
    "title": "Speaker-Consistent Grounding for Confusion-Resilient LLM-Based Target Speech Extraction",
    "snapshot_date": "2026-10-25",
    "description": "Transcribed from Tables 1 and 2 of the current author-provided manuscript. Values retain the manuscript's displayed precision.",
    "table_1_page": 3,
    "table_2_page": 4,
    "pdf_sha256": "33d2e4670a223b13b468ae5856a35ca443386748c0198bc6aeb8d7246147165b"
  },
  "protocol": {
    "clean_test_trials": 6000,
    "noisy_test_trials": 8400,
    "noisy_natural_trials": 6000,
    "noisy_controlled_snr_trials": 2400,
    "controlled_trials_per_snr": 480,
    "controlled_snr_db": [
      -5,
      0,
      5,
      10,
      15
    ],
    "wer": "Uncapped macro WER against Whisper-large-v3 transcripts of the clean target, not human transcripts.",
    "acoustic_switch": "Negative target-minus-interferer CAMPPlus cosine-similarity margin.",
    "wrong_speaker_error": "Negative clean-reference SI-SDR margin; distinct from acoustic-switch rate.",
    "quality": "DNSMOS P.808 and UTMOS are predicted quality scores, not listening-test MOS."
  },
  "columns": [
    "wer_percent",
    "content_switch_percent",
    "acoustic_switch_percent",
    "speaker_margin",
    "dnsmos_p808",
    "utmos"
  ],
  "systems": [
    "BSRNN-TSE",
    "BSRNN-TSE-CDCS-5",
    "Qwen-TSE-UD",
    "Qwen-TSE-fixed-CSG",
    "Qwen-TSE-adaptive-CSG"
  ],
  "clean_test": [
    [
      11.41,
      7.0,
      6.98,
      0.449,
      3.688,
      3.606
    ],
    [
      7.02,
      0.88,
      0.78,
      0.506,
      3.699,
      3.647
    ],
    [
      8.99,
      0.92,
      0.43,
      0.421,
      3.705,
      3.913
    ],
    [
      8.04,
      0.92,
      0.23,
      0.422,
      3.699,
      3.91
    ],
    [
      8.12,
      0.9,
      0.2,
      0.421,
      3.711,
      3.921
    ]
  ],
  "noisy_test": [
    [
      39.85,
      18.79,
      18.23,
      0.266,
      3.031,
      2.001
    ],
    [
      30.67,
      11.02,
      10.39,
      0.327,
      2.98,
      1.948
    ],
    [
      46.7,
      12.79,
      1.68,
      0.377,
      3.494,
      3.216
    ],
    [
      42.11,
      9.1,
      0.51,
      0.386,
      3.498,
      3.147
    ],
    [
      42.73,
      9.37,
      0.53,
      0.388,
      3.503,
      3.253
    ]
  ],
  "ablation_columns": [
    "clean_dev",
    "clean_test",
    "noisy_dev",
    "noisy_test"
  ],
  "ablation": [
    {
      "configuration": "Full (no CDCS)",
      "values": [
        6.9,
        6.87,
        19.85,
        18.95
      ]
    },
    {
      "configuration": "Early (E)",
      "values": [
        8.52,
        8.57,
        20.61,
        19.96
      ]
    },
    {
      "configuration": "Middle (M)",
      "values": [
        8.27,
        8.9,
        20.92,
        20.73
      ]
    },
    {
      "configuration": "Late (L)",
      "values": [
        8.52,
        8.5,
        20.71,
        20.18
      ]
    },
    {
      "configuration": "TF-context (TF)",
      "values": [
        2.35,
        2.53,
        14.32,
        14.37
      ]
    },
    {
      "configuration": "Full + E/M/L",
      "values": [
        3.7,
        3.6,
        15.96,
        15.0
      ]
    },
    {
      "configuration": "Full + TF",
      "values": [
        1.17,
        1.18,
        12.0,
        11.3
      ]
    },
    {
      "configuration": "CDCS-5",
      "values": [
        0.88,
        0.75,
        11.02,
        10.46
      ]
    }
  ],
  "recovery": [
    {
      "split": "Clean TEST",
      "recovery": "355/398 (89.20%)",
      "oracle": "360/398 (90.45%)",
      "new_errors": "2/5,588 (0.04%)"
    },
    {
      "split": "Noisy TEST",
      "recovery": "374/646 (57.89%)",
      "oracle": "431/646 (66.72%)",
      "new_errors": "8/4,709 (0.17%)"
    }
  ]
}