{
  "schema_version": 4,
  "title": "Natural Noisy TEST LLM-WER listening case study — Whisper large-v3",
  "scope": {
    "split": "Natural Noisy TEST",
    "trial_count": 6000,
    "post_hoc": true,
    "test_used": true,
    "configuration_selected_from_test": false,
    "does_not_replace_frozen_test_metrics": true,
    "human_adjudication": false,
    "interpretation": "Illustrative listening cases, not prevalence estimates. High WER alone is not labelled as an audible waveform error.",
    "audit_completion": {
      "unique_transcripts": 57578,
      "metrics_rows": 50400,
      "systems": 6,
      "trials_per_system": 8400,
      "natural_trials_per_system": 6000,
      "failures": 0,
      "merge_validation": "PASS"
    },
    "audio_terms": "See AUDIO_ATTRIBUTION.md.",
    "asr": {
      "model": "openai/whisper-large-v3",
      "revision": "06f233fe06e710322aca913c1bc4249a0d71fce1",
      "task": "English/transcribe",
      "decoding": "greedy",
      "repetition_penalty": 1.1,
      "batch_size": 32
    },
    "reference_protocol": "Config-consistent WER: each system output is scored against the clean-target transcript produced by the same Whisper large-v3 configuration."
  },
  "cases": [
    {
      "id": "grounding-repair-ud",
      "title": "Grounding removes residual UD lexical drift",
      "failure_mode": "UD lexical drift repaired by grounding",
      "selection_rule": "Illustrative complete Natural Noisy TEST trial in which large-v3 measures 26.67% WER for UD and 0% for fixed, adaptive, and GNR grounding.",
      "listening_note": "The UD waveform remains intelligible but lexically imperfect; compare it with the target-exact grounded outputs.",
      "target_text": "he was like unto my father in a way and yet was not my father",
      "systems": {
        "bsrnn_tse_cdcs_5": {
          "label": "BSRNN-TSE-CDCS-5",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.614222,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/grounding-repair-ud/cdcs-5.wav",
            "sha256": "1d5aa02de6ea8b212b41de517962decb2ac83287b53bd9312daf1ae697a47ca9",
            "sample_rate_hz": 16000,
            "duration_seconds": 4.312
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 0,
            "substitutions": 0,
            "deletions": 0,
            "insertions": 0,
            "reference_words": 15,
            "text": "he was like unto my father in a way and yet was not my father",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_ud": {
          "label": "Qwen-TSE-UD",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.472376,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/grounding-repair-ud/qwen-ud.wav",
            "sha256": "b7ac3eb494eb2f4bc10c9e5c60162596710f9625f5fb7b67b1ae7fd37a563f89",
            "sample_rate_hz": 24000,
            "duration_seconds": 4.32
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 26.666667,
            "substitutions": 3,
            "deletions": 0,
            "insertions": 1,
            "reference_words": 15,
            "text": "she was reckoned to be my father in a way and yet was not my father",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_fixed_csg": {
          "label": "Qwen-TSE-fixed-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.457797,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/grounding-repair-ud/qwen-fixed-csg.wav",
            "sha256": "7b338f86410a8a32aa0d65b1a3bfb4b458933a762af195bd7e32c072d90a089c",
            "sample_rate_hz": 24000,
            "duration_seconds": 4.32
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 0,
            "substitutions": 0,
            "deletions": 0,
            "insertions": 0,
            "reference_words": 15,
            "text": "he was like unto my father in a way and yet was not my father",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg": {
          "label": "Qwen-TSE-adaptive-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.46717,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/grounding-repair-ud/qwen-adaptive-csg.wav",
            "sha256": "5776a99cdfbf64953b9841e670310586504634785132dc4af1144dbb63add3e0",
            "sample_rate_hz": 24000,
            "duration_seconds": 4.32
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 0,
            "substitutions": 0,
            "deletions": 0,
            "insertions": 0,
            "reference_words": 15,
            "text": "he was like unto my father in a way and yet was not my father",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg_gnr": {
          "label": "Qwen-TSE-adaptive-CSG+GNR (supplementary)",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.494469,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/grounding-repair-ud/qwen-adaptive-csg-gnr.wav",
            "sha256": "ef30385afd6cf23c76c267cf9db17316f300665c7a1231ee631ff7524e7740fd",
            "sample_rate_hz": 24000,
            "duration_seconds": 4.32
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 0,
            "substitutions": 0,
            "deletions": 0,
            "insertions": 0,
            "reference_words": 15,
            "text": "he was like unto my father in a way and yet was not my father",
            "hard_limit_proxy": false
          }
        }
      },
      "audio": {
        "mixture": {
          "path": "public/assets/audio/noisy-llm-case-study/grounding-repair-ud/mixture.wav",
          "sha256": "17782ea8be0b731ed63b8d5f2c90b222d5a27da376c119ab8c2800b8740219ba",
          "sample_rate_hz": 16000,
          "duration_seconds": 4.315
        },
        "target": {
          "path": "public/assets/audio/noisy-llm-case-study/grounding-repair-ud/target.wav",
          "sha256": "7140c034ff9edb842c22f9ef5850bd006835aa27c21fc9210b856edc07bea6cc",
          "sample_rate_hz": 16000,
          "duration_seconds": 4.315
        },
        "interferer": {
          "path": "public/assets/audio/noisy-llm-case-study/grounding-repair-ud/interferer.wav",
          "sha256": "d786a435a82cf1205f6098e8ca0c1fd1507de8414a0b09b37793ee8bb9c796fe",
          "sample_rate_hz": 16000,
          "duration_seconds": 4.315
        }
      }
    },
    {
      "id": "largev3-runaway",
      "title": "Whisper large-v3 can enter a decoding loop",
      "failure_mode": "Large-v3 decoding runaway",
      "selection_rule": "Largest large-v3 insertion count among complete Natural Noisy TEST outputs.",
      "listening_note": "Large-v3 inserts 419 repeated tokens. Listen to the complete waveform to judge whether that repetition is acoustic or measurement-side.",
      "target_text": "alexander paced up and down the hallway buttoning and unbuttoning his overcoat until she returned and took him up to hildas living room",
      "systems": {
        "bsrnn_tse_cdcs_5": {
          "label": "BSRNN-TSE-CDCS-5",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.547999,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/largev3-runaway/cdcs-5.wav",
            "sha256": "6d5d8100f82d0a11ca6b42cc242cce7ce44caa0022213d68ac3ad9c645f801a4",
            "sample_rate_hz": 16000,
            "duration_seconds": 7.672
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 8.695652,
            "substitutions": 1,
            "deletions": 1,
            "insertions": 0,
            "reference_words": 23,
            "text": "alexander paced up and down the hallway buttoning and unbuttoning his overcoat until she returned and took off to hildas living room",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_ud": {
          "label": "Qwen-TSE-UD",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.466915,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/largev3-runaway/qwen-ud.wav",
            "sha256": "b19c1c7080b94ff075f11d81bdba91627c07b9cf548134d3ceceb36d8d6a755d",
            "sample_rate_hz": 24000,
            "duration_seconds": 7.68
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 1895.652174,
            "substitutions": 17,
            "deletions": 0,
            "insertions": 419,
            "reference_words": 23,
            "text": "out zander paced up and down the hallway f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_fixed_csg": {
          "label": "Qwen-TSE-fixed-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.485288,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/largev3-runaway/qwen-fixed-csg.wav",
            "sha256": "725d1971ddfb47ae891447a5f06d62264bcfc6f1fe0cad8133f1cfeb8999fe1d",
            "sample_rate_hz": 24000,
            "duration_seconds": 7.68
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 34.782609,
            "substitutions": 5,
            "deletions": 3,
            "insertions": 0,
            "reference_words": 23,
            "text": "alexander paced up and down the hallway buttoning and unbuttoning his overcoat until she returned into the hotel of his",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg": {
          "label": "Qwen-TSE-adaptive-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.477791,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/largev3-runaway/qwen-adaptive-csg.wav",
            "sha256": "8753a6aa73968efbf21ba6f88a3fbad7c054c9217186e1a20852f30b9e6c61f6",
            "sample_rate_hz": 24000,
            "duration_seconds": 7.68
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 34.782609,
            "substitutions": 5,
            "deletions": 3,
            "insertions": 0,
            "reference_words": 23,
            "text": "alexander paced up and down the hallway buttoning and unbuttoning his overcoat until she returned into the tail of his",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg_gnr": {
          "label": "Qwen-TSE-adaptive-CSG+GNR (supplementary)",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.483429,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/largev3-runaway/qwen-adaptive-csg-gnr.wav",
            "sha256": "6a771cc38da272ff7d8b58d592003a489ec53d424140c3f9b647a0f5498f9c32",
            "sample_rate_hz": 24000,
            "duration_seconds": 7.68
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 34.782609,
            "substitutions": 5,
            "deletions": 3,
            "insertions": 0,
            "reference_words": 23,
            "text": "alexander paced up and down the hallway buttoning and unbuttoning his overcoat until she returned into the tail of his",
            "hard_limit_proxy": false
          }
        }
      },
      "audio": {
        "mixture": {
          "path": "public/assets/audio/noisy-llm-case-study/largev3-runaway/mixture.wav",
          "sha256": "a15dd5813807acc1393041d4a44600e0d5073524a7638741d68687371a9130a0",
          "sample_rate_hz": 16000,
          "duration_seconds": 7.675
        },
        "target": {
          "path": "public/assets/audio/noisy-llm-case-study/largev3-runaway/target.wav",
          "sha256": "40c3c6b8668c5dc0b329cb9102535c0c20da2ef253059cb20fbe3e7139a11c66",
          "sample_rate_hz": 16000,
          "duration_seconds": 7.675
        },
        "interferer": {
          "path": "public/assets/audio/noisy-llm-case-study/largev3-runaway/interferer.wav",
          "sha256": "f24ece1ee60dc637032e3c070b3cc42066e57350a643e762868feaa427de8c4c",
          "sample_rate_hz": 16000,
          "duration_seconds": 7.675
        }
      }
    },
    {
      "id": "ud-collapse",
      "title": "Large-v3 hears no lexical content from UD",
      "failure_mode": "UD waveform-content collapse",
      "selection_rule": "Complete Natural Noisy TEST trial with an empty large-v3 UD transcript and 100% deletion WER.",
      "listening_note": "The UD file has normal nominal duration, but large-v3 returns an empty transcript. Listen for silence, nonspeech, or speech too degraded to recover.",
      "target_text": "steam up and canvas spread the schooner started eastwards",
      "systems": {
        "bsrnn_tse_cdcs_5": {
          "label": "BSRNN-TSE-CDCS-5",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.442696,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/ud-collapse/cdcs-5.wav",
            "sha256": "da925367ddb469048caf9a095e33f308a27bc78de9c9e6fd2f379d762804dfa3",
            "sample_rate_hz": 16000,
            "duration_seconds": 3.976
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 55.555556,
            "substitutions": 4,
            "deletions": 0,
            "insertions": 1,
            "reference_words": 9,
            "text": "theme of the canvas spread the schooner started to lose",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_ud": {
          "label": "Qwen-TSE-UD",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.057863,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/ud-collapse/qwen-ud.wav",
            "sha256": "cc619cd20e12b0393f07fa7d8eb6078a814ed185d4c50afd2f99d877ddc03bf4",
            "sample_rate_hz": 24000,
            "duration_seconds": 4
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 100,
            "substitutions": 0,
            "deletions": 9,
            "insertions": 0,
            "reference_words": 9,
            "text": "",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_fixed_csg": {
          "label": "Qwen-TSE-fixed-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.534271,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/ud-collapse/qwen-fixed-csg.wav",
            "sha256": "7fa3240f2c08b8d1550d73535a2057a9e6c6ed968145d8f7cf9f808cae8d3f0a",
            "sample_rate_hz": 24000,
            "duration_seconds": 4
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 66.666667,
            "substitutions": 4,
            "deletions": 2,
            "insertions": 0,
            "reference_words": 9,
            "text": "theba cant spread the schooner starting to",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg": {
          "label": "Qwen-TSE-adaptive-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.525914,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/ud-collapse/qwen-adaptive-csg.wav",
            "sha256": "9b20ed462ff44cb003a56dc8b60d60825f4124530ccc03796c0480a1df2d5373",
            "sample_rate_hz": 24000,
            "duration_seconds": 4
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 88.888889,
            "substitutions": 5,
            "deletions": 1,
            "insertions": 2,
            "reference_words": 9,
            "text": "thebop cant let spread the schooner start in the woods",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg_gnr": {
          "label": "Qwen-TSE-adaptive-CSG+GNR (supplementary)",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.528047,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/ud-collapse/qwen-adaptive-csg-gnr.wav",
            "sha256": "0797713a8fd0efe5e498153f2dd5dc41fc3488226fe45e1323f8b3da797fbfa3",
            "sample_rate_hz": 24000,
            "duration_seconds": 4
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 100,
            "substitutions": 8,
            "deletions": 0,
            "insertions": 1,
            "reference_words": 9,
            "text": "thebop cant whisper and the schooner starts in the woods",
            "hard_limit_proxy": false
          }
        }
      },
      "audio": {
        "mixture": {
          "path": "public/assets/audio/noisy-llm-case-study/ud-collapse/mixture.wav",
          "sha256": "3e72a5118f4d9cdad9f97054719ea24eaedf1390719525bb08c1df4d3dcca3b7",
          "sample_rate_hz": 16000,
          "duration_seconds": 3.98
        },
        "target": {
          "path": "public/assets/audio/noisy-llm-case-study/ud-collapse/target.wav",
          "sha256": "6d51e6815d01c32315c55130077ecadb00863a1459e507b5b5e2b49e6208d035",
          "sample_rate_hz": 16000,
          "duration_seconds": 3.98
        },
        "interferer": {
          "path": "public/assets/audio/noisy-llm-case-study/ud-collapse/interferer.wav",
          "sha256": "dc7b2105123608bdda4723afd1a94a8c4395ad86f62e8fb38954d7f84775cfd5",
          "sample_rate_hz": 16000,
          "duration_seconds": 3.98
        }
      }
    },
    {
      "id": "identity-content-drift",
      "title": "Target-like voice, target-inconsistent words",
      "failure_mode": "Acoustic identity/content separation",
      "selection_rule": "High config-consistent large-v3 WER across all LLM arms while every LLM arm retains a positive speaker margin and no acoustic speaker switch.",
      "listening_note": "The output remains target-like according to the speaker encoder, but its fluent sentence is far from the 25-word large-v3 target reference.",
      "target_text": "ill gladly do that promised the new boolaroo and ill feed the honourable goat all the shavings and leather and tin cans he can eat",
      "systems": {
        "bsrnn_tse_cdcs_5": {
          "label": "BSRNN-TSE-CDCS-5",
          "acoustic_speaker_switch": true,
          "speaker_margin": -0.317458,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/identity-content-drift/cdcs-5.wav",
            "sha256": "a12a5d2013e93af931574f3833a48cc4d4d5edadcd30830dadef01e38779f2eb",
            "sample_rate_hz": 16000,
            "duration_seconds": 9
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 104,
            "substitutions": 23,
            "deletions": 0,
            "insertions": 3,
            "reference_words": 25,
            "text": "of course he did not realize that it was i who was knocking but nonetheless his conduct was very uncourteous and indeed under the circumstances rather suspicious sigh",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_ud": {
          "label": "Qwen-TSE-UD",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.107256,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/identity-content-drift/qwen-ud.wav",
            "sha256": "4a2910c1d0c48f8c32a7a06ad8534a3eeb5511bbeb04d176ff46762c7842b375",
            "sample_rate_hz": 24000,
            "duration_seconds": 9.04
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 124,
            "substitutions": 23,
            "deletions": 0,
            "insertions": 8,
            "reference_words": 25,
            "text": "of course he did not realize that it was i who was nodding but none the less his conduct was very uncourteous i need a deed out of circumstance drillers suspicion was fine",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_fixed_csg": {
          "label": "Qwen-TSE-fixed-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.137775,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/identity-content-drift/qwen-fixed-csg.wav",
            "sha256": "ce77c1331a769251c423ee0bbddef9ba00a4e6e3c49e93e63c9205e4df9fca07",
            "sample_rate_hz": 24000,
            "duration_seconds": 9.04
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 112,
            "substitutions": 22,
            "deletions": 0,
            "insertions": 6,
            "reference_words": 25,
            "text": "of course he did not realize that it was i who was nodding but nonetheless his conduct was very uncourteous and indeed out of the circumstance he was rather suspicious fine",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg": {
          "label": "Qwen-TSE-adaptive-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.132823,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/identity-content-drift/qwen-adaptive-csg.wav",
            "sha256": "fe6906d2e712b273830723a6a3eed20bd35df77c5c4736dd25178bfa95227a3d",
            "sample_rate_hz": 24000,
            "duration_seconds": 9.04
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 112,
            "substitutions": 22,
            "deletions": 0,
            "insertions": 6,
            "reference_words": 25,
            "text": "of course he did not realize that it was i who was nodding but nonetheless his conduct was very uncourteous and indeed out of the circumstance he was rather suspicious fine",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg_gnr": {
          "label": "Qwen-TSE-adaptive-CSG+GNR (supplementary)",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.132823,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/identity-content-drift/qwen-adaptive-csg-gnr.wav",
            "sha256": "fe6906d2e712b273830723a6a3eed20bd35df77c5c4736dd25178bfa95227a3d",
            "sample_rate_hz": 24000,
            "duration_seconds": 9.04
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 112,
            "substitutions": 22,
            "deletions": 0,
            "insertions": 6,
            "reference_words": 25,
            "text": "of course he did not realize that it was i who was nodding but nonetheless his conduct was very uncourteous and indeed out of the circumstance he was rather suspicious fine",
            "hard_limit_proxy": false
          }
        }
      },
      "audio": {
        "mixture": {
          "path": "public/assets/audio/noisy-llm-case-study/identity-content-drift/mixture.wav",
          "sha256": "ca3af145ce01799a953c13b80b4d1b04fb9240529f059c279c8cece450b76bb0",
          "sample_rate_hz": 16000,
          "duration_seconds": 9.005
        },
        "target": {
          "path": "public/assets/audio/noisy-llm-case-study/identity-content-drift/target.wav",
          "sha256": "13fe1b2c0225b12580f5e28a59cf2ecb7463b0938016ab8d95d3e0fb87057050",
          "sample_rate_hz": 16000,
          "duration_seconds": 9.005
        },
        "interferer": {
          "path": "public/assets/audio/noisy-llm-case-study/identity-content-drift/interferer.wav",
          "sha256": "1236beb97452acf60f70ddbdaba852cf20995dae8f0d71aaf5411a3edebcbd00",
          "sample_rate_hz": 16000,
          "duration_seconds": 9.005
        }
      }
    },
    {
      "id": "fixed-csg-repair",
      "title": "Fixed CSG repairs an UD identity/content collapse",
      "failure_mode": "Grounding repair",
      "selection_rule": "UD has 100% config-consistent large-v3 WER and a negative speaker margin, while fixed CSG is exact and restores a positive margin.",
      "listening_note": "Large-v3 transcribes UD as ‘thank you’; fixed CSG restores the complete target sentence and target-side speaker identity.",
      "target_text": "suddenly the ichthyosaurus and the plesiosaurus disappear below",
      "systems": {
        "bsrnn_tse_cdcs_5": {
          "label": "BSRNN-TSE-CDCS-5",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.621726,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-repair/cdcs-5.wav",
            "sha256": "8fd9211f6c05e2233c99917d7f5d5eb59e4ce64658dc20d3a530ebe5cff689d6",
            "sample_rate_hz": 16000,
            "duration_seconds": 5.12
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 0,
            "substitutions": 0,
            "deletions": 0,
            "insertions": 0,
            "reference_words": 8,
            "text": "suddenly the ichthyosaurus and the plesiosaurus disappear below",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_ud": {
          "label": "Qwen-TSE-UD",
          "acoustic_speaker_switch": true,
          "speaker_margin": -0.03911,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-repair/qwen-ud.wav",
            "sha256": "21e9b86006dfed8ed9b57b8a2fa8beb605a65d8a1683eab08ed84373bd9aad0c",
            "sample_rate_hz": 24000,
            "duration_seconds": 5.16
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 100,
            "substitutions": 2,
            "deletions": 6,
            "insertions": 0,
            "reference_words": 8,
            "text": "thank you",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_fixed_csg": {
          "label": "Qwen-TSE-fixed-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.396871,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-repair/qwen-fixed-csg.wav",
            "sha256": "0ff592779cb5055023d2bdd5a17790e1663a019b47a86078439df8b21cba1482",
            "sample_rate_hz": 24000,
            "duration_seconds": 5.16
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 0,
            "substitutions": 0,
            "deletions": 0,
            "insertions": 0,
            "reference_words": 8,
            "text": "suddenly the ichthyosaurus and the plesiosaurus disappear below",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg": {
          "label": "Qwen-TSE-adaptive-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.401422,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-repair/qwen-adaptive-csg.wav",
            "sha256": "5b9e052828c3fb942a80c14764d4f88b2dd3e28ad86303fe76d7cb8cd8605c43",
            "sample_rate_hz": 24000,
            "duration_seconds": 5.16
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 0,
            "substitutions": 0,
            "deletions": 0,
            "insertions": 0,
            "reference_words": 8,
            "text": "suddenly the ichthyosaurus and the plesiosaurus disappear below",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg_gnr": {
          "label": "Qwen-TSE-adaptive-CSG+GNR (supplementary)",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.356903,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-repair/qwen-adaptive-csg-gnr.wav",
            "sha256": "b2235458628490cafe3bee34207f9d804300353da975647a46f589d8ce611551",
            "sample_rate_hz": 24000,
            "duration_seconds": 5.16
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 0,
            "substitutions": 0,
            "deletions": 0,
            "insertions": 0,
            "reference_words": 8,
            "text": "suddenly the ichthyosaurus and the plesiosaurus disappear below",
            "hard_limit_proxy": false
          }
        }
      },
      "audio": {
        "mixture": {
          "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-repair/mixture.wav",
          "sha256": "3ff48c8023b87390e7c72b3f3040f1f0d796b4bcab050e440867d86ef809d6eb",
          "sample_rate_hz": 16000,
          "duration_seconds": 5.125
        },
        "target": {
          "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-repair/target.wav",
          "sha256": "ad5b658d3b7eb4a42cfc317db046ef7e927c66f9ae23123ef8483c7e543e8774",
          "sample_rate_hz": 16000,
          "duration_seconds": 5.125
        },
        "interferer": {
          "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-repair/interferer.wav",
          "sha256": "e970be34153d673b633f8189851564657c8599403158e3b0c27224efe57b6fe2",
          "sample_rate_hz": 16000,
          "duration_seconds": 5.125
        }
      }
    },
    {
      "id": "fixed-csg-regression",
      "title": "A counterexample: fixed CSG invents fluent unrelated speech",
      "failure_mode": "Grounding regression",
      "selection_rule": "Natural Noisy TEST regression where fixed CSG worsens config-consistent large-v3 WER and introduces an acoustic speaker switch.",
      "listening_note": "UD is nearly empty, but fixed CSG replaces it with fluent unrelated content and moves the acoustic speaker margin below zero.",
      "target_text": "sometime youll tell me please wont you",
      "systems": {
        "bsrnn_tse_cdcs_5": {
          "label": "BSRNN-TSE-CDCS-5",
          "acoustic_speaker_switch": true,
          "speaker_margin": -0.083588,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-regression/cdcs-5.wav",
            "sha256": "3bc13ce77c0b2ae463e531febe568cdc60e606d2baa562383f08b2e63af88162",
            "sample_rate_hz": 16000,
            "duration_seconds": 3.288
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 214.285714,
            "substitutions": 7,
            "deletions": 0,
            "insertions": 8,
            "reference_words": 7,
            "text": "i used to have an advantage on putting them together but now im a chico",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_ud": {
          "label": "Qwen-TSE-UD",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.033917,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-regression/qwen-ud.wav",
            "sha256": "a8158d190c3b8e396d3a3b1f7abe18ec631794b84bccf840bf10d9539504cbf1",
            "sample_rate_hz": 24000,
            "duration_seconds": 3.32
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 85.714286,
            "substitutions": 0,
            "deletions": 6,
            "insertions": 0,
            "reference_words": 7,
            "text": "you",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_fixed_csg": {
          "label": "Qwen-TSE-fixed-CSG",
          "acoustic_speaker_switch": true,
          "speaker_margin": -0.035985,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-regression/qwen-fixed-csg.wav",
            "sha256": "172fdaf8d09c06256b8f8519a55a5f4779aefeb8a32eba43355fdf19f9053308",
            "sample_rate_hz": 24000,
            "duration_seconds": 3.32
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 128.571429,
            "substitutions": 7,
            "deletions": 0,
            "insertions": 2,
            "reference_words": 7,
            "text": "for me i was just like oh my god",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg": {
          "label": "Qwen-TSE-adaptive-CSG",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.009156,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-regression/qwen-adaptive-csg.wav",
            "sha256": "85b4c609cd6c11d40b30caeced20c1bd2cd1f4bcbe5e36bb60a6d12138dece01",
            "sample_rate_hz": 24000,
            "duration_seconds": 3.32
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 185.714286,
            "substitutions": 7,
            "deletions": 0,
            "insertions": 6,
            "reference_words": 7,
            "text": "for me it was the yeah it was a gap oh my chicken",
            "hard_limit_proxy": false
          }
        },
        "qwen_tse_adaptive_csg_gnr": {
          "label": "Qwen-TSE-adaptive-CSG+GNR (supplementary)",
          "acoustic_speaker_switch": false,
          "speaker_margin": 0.0027,
          "audio": {
            "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-regression/qwen-adaptive-csg-gnr.wav",
            "sha256": "1517658747b66ca9325fd8ddb27683b317b30eef989bb23bcfa8174f2109dec7",
            "sample_rate_hz": 24000,
            "duration_seconds": 3.32
          },
          "whisper_large_v3_rp1p10": {
            "wer_percent": 185.714286,
            "substitutions": 7,
            "deletions": 0,
            "insertions": 6,
            "reference_words": 7,
            "text": "for me it was just a little music id be like oh chicken",
            "hard_limit_proxy": false
          }
        }
      },
      "audio": {
        "mixture": {
          "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-regression/mixture.wav",
          "sha256": "ceccf7d159e4875afc234312ade89a6d347a1b454c35c23a380167eb4f81553e",
          "sample_rate_hz": 16000,
          "duration_seconds": 3.295
        },
        "target": {
          "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-regression/target.wav",
          "sha256": "afb6c01a06fd3beb27b307f809109aab80a78c6eba26bcb249afad6d458edd83",
          "sample_rate_hz": 16000,
          "duration_seconds": 3.295
        },
        "interferer": {
          "path": "public/assets/audio/noisy-llm-case-study/fixed-csg-regression/interferer.wav",
          "sha256": "8b00c1e3c5dccea518796ce8a4150f13047efd00bf97fcbde579d10283dadb7e",
          "sample_rate_hz": 16000,
          "duration_seconds": 3.295
        }
      }
    }
  ]
}