{ "analysis": "trailing_silence_sensitivity", "checkpoint": "artifacts/partial-shard-warmstart-lr3e4-5ep/best.pt", "checkpoint_sha256": "e919cdbdc89046bd82fccf52a460406643e917b06d42a966d0e0104b2974f6c1", "conditions": [ { "classification_metrics": { "accuracy": 0.5306748466257669, "average_precision": 0.6880689959324726, "balanced_accuracy": 0.5222138554216867, "brier_score": 0.2133037066638398, "count": 326, "expected_calibration_error": 0.1055241439925747, "f1": 0.11560693641618497, "false_negative_rate": 0.9375, "false_positive_rate": 0.018072289156626505, "fn": 150, "fp": 3, "log_loss": 0.6135001064370812, "macro_f1": 0.39809574378220525, "negative_count": 166, "positive_count": 160, "precision": 0.7692307692307693, "recall": 0.0625, "roc_auc": 0.7389683734939759, "specificity": 0.9819277108433735, "threshold": 0.741000771522522, "tn": 163, "tp": 10 }, "probability_summary": { "count": 326, "maximum": 0.7886796593666077, "mean": 0.4128557691133461, "minimum": 0.0803893655538559, "p05": 0.11172053776681423, "p50": 0.4176659733057022, "p95": 0.7287024110555649, "standard_deviation": 0.2055378774150228 }, "relative_to_0ms": null, "trailing_silence_ms": 0 }, { "classification_metrics": { "accuracy": 0.5276073619631901, "average_precision": 0.6840183571897368, "balanced_accuracy": 0.5193147590361445, "brier_score": 0.2100618660450317, "count": 326, "expected_calibration_error": 0.10356104012830125, "f1": 0.125, "false_negative_rate": 0.93125, "false_positive_rate": 0.030120481927710843, "fn": 149, "fp": 5, "log_loss": 0.605744453547225, "macro_f1": 0.4007352941176471, "negative_count": 166, "positive_count": 160, "precision": 0.6875, "recall": 0.06875, "roc_auc": 0.7403237951807229, "specificity": 0.9698795180722891, "threshold": 0.741000771522522, "tn": 161, "tp": 11 }, "probability_summary": { "count": 326, "maximum": 0.7887353301048279, "mean": 0.4271308436898366, "minimum": 0.07923124730587006, "p05": 0.10842713713645935, "p50": 0.44254499673843384, "p95": 0.7391214519739151, "standard_deviation": 0.21076642495080192 }, "relative_to_0ms": { "probability_shift": { "by_label": { "END": { "count": 160, "maximum": 0.4578339457511902, "mean": 0.01773580485023558, "minimum": -0.07176662981510162, "p05": -0.02797318696975708, "p50": 0.00010977685451507568, "p95": 0.10832937657833029, "standard_deviation": 0.06593732941800336 }, "HOLD": { "count": 166, "maximum": 0.310523584485054, "mean": 0.010939430939145836, "minimum": -0.08172200620174408, "p05": -0.030832231044769287, "p50": -0.002315342426300049, "p95": 0.1006043553352356, "standard_deviation": 0.05458234250596568 } }, "count": 326, "decreased_count": 181, "definition": "p_end(condition) - p_end(0ms) on the same examples", "increased_count": 145, "maximum": 0.4578339457511902, "mean": 0.014275074576490496, "mean_absolute": 0.02900380597798371, "minimum": -0.08172200620174408, "p05": -0.029261156916618347, "p50": -0.001150473952293396, "p95": 0.10375719517469406, "root_mean_square": 0.062178852734470415, "standard_deviation": 0.06051803014978754, "unchanged_count": 0 }, "threshold_decision_flips": { "END_to_HOLD": 1, "HOLD_to_END": 4, "count": 5, "false_interruptions_introduced": 2, "false_interruptions_resolved": 0, "missed_ends_introduced": 1, "missed_ends_resolved": 2, "rate": 0.015337423312883436, "unchanged_count": 321 } }, "trailing_silence_ms": 200 }, { "classification_metrics": { "accuracy": 0.5337423312883436, "average_precision": 0.6979212626611389, "balanced_accuracy": 0.5255647590361445, "brier_score": 0.2072492112934809, "count": 326, "expected_calibration_error": 0.06585103341955346, "f1": 0.14606741573033707, "false_negative_rate": 0.91875, "false_positive_rate": 0.030120481927710843, "fn": 147, "fp": 5, "log_loss": 0.598821131080406, "macro_f1": 0.41269615512255253, "negative_count": 166, "positive_count": 160, "precision": 0.7222222222222222, "recall": 0.08125, "roc_auc": 0.7427710843373494, "specificity": 0.9698795180722891, "threshold": 0.741000771522522, "tn": 161, "tp": 13 }, "probability_summary": { "count": 326, "maximum": 0.7873576879501343, "mean": 0.4418996488846884, "minimum": 0.07847821712493896, "p05": 0.10653392039239407, "p50": 0.4775547683238983, "p95": 0.7417303323745728, "standard_deviation": 0.2172591006725531 }, "relative_to_0ms": { "probability_shift": { "by_label": { "END": { "count": 160, "maximum": 0.5757421255111694, "mean": 0.03650752436369657, "minimum": -0.26492756605148315, "p05": -0.06253614425659179, "p50": 0.0008754730224609375, "p95": 0.27894054204225516, "standard_deviation": 0.11404161626592267 }, "HOLD": { "count": 166, "maximum": 0.5056779384613037, "mean": 0.021850005465458674, "minimum": -0.15796267986297607, "p05": -0.06120702624320984, "p50": -0.004482690244913101, "p95": 0.1999836601316929, "standard_deviation": 0.09575172708096644 } }, "count": 326, "decreased_count": 181, "definition": "p_end(condition) - p_end(0ms) on the same examples", "increased_count": 145, "maximum": 0.5757421255111694, "mean": 0.029043879771342308, "mean_absolute": 0.05793183815716966, "minimum": -0.26492756605148315, "p05": -0.06249040365219116, "p50": -0.002118140459060669, "p95": 0.2521081864833832, "root_mean_square": 0.10931088923065266, "standard_deviation": 0.10538179896084442, "unchanged_count": 0 }, "threshold_decision_flips": { "END_to_HOLD": 4, "HOLD_to_END": 9, "count": 13, "false_interruptions_introduced": 3, "false_interruptions_resolved": 1, "missed_ends_introduced": 3, "missed_ends_resolved": 6, "rate": 0.03987730061349693, "unchanged_count": 313 } }, "trailing_silence_ms": 400 }, { "classification_metrics": { "accuracy": 0.5613496932515337, "average_precision": 0.7463143655056782, "balanced_accuracy": 0.5538027108433735, "brier_score": 0.19892348061598683, "count": 326, "expected_calibration_error": 0.07515490173927845, "f1": 0.24338624338624337, "false_negative_rate": 0.85625, "false_positive_rate": 0.03614457831325301, "fn": 137, "fp": 6, "log_loss": 0.5786536033097901, "macro_f1": 0.46726547590478473, "negative_count": 166, "positive_count": 160, "precision": 0.7931034482758621, "recall": 0.14375, "roc_auc": 0.7672063253012048, "specificity": 0.963855421686747, "threshold": 0.741000771522522, "tn": 160, "tp": 23 }, "probability_summary": { "count": 326, "maximum": 0.8118427395820618, "mean": 0.46326974340560245, "minimum": 0.07712869346141815, "p05": 0.09877880848944187, "p50": 0.5251197218894958, "p95": 0.7567593455314636, "standard_deviation": 0.2247464959572825 }, "relative_to_0ms": { "probability_shift": { "by_label": { "END": { "count": 160, "maximum": 0.643292024731636, "mean": 0.06806713282130658, "minimum": -0.28288304805755615, "p05": -0.1087776966392994, "p50": 0.006056338548660278, "p95": 0.4468590237200257, "standard_deviation": 0.1681885565157826 }, "HOLD": { "count": 166, "maximum": 0.5687761902809143, "mean": 0.03339888173413564, "minimum": -0.2930072098970413, "p05": -0.12552335858345032, "p50": -0.015010792762041092, "p95": 0.3289465010166168, "standard_deviation": 0.14780063874515106 } }, "count": 326, "decreased_count": 178, "definition": "p_end(condition) - p_end(0ms) on the same examples", "increased_count": 148, "maximum": 0.643292024731636, "mean": 0.05041397429225635, "mean_absolute": 0.10576534182107522, "minimum": -0.2930072098970413, "p05": -0.12200940027832985, "p50": -0.0059654973447322845, "p95": 0.42330726981163025, "root_mean_square": 0.16687978628652683, "standard_deviation": 0.15908266488557635, "unchanged_count": 0 }, "threshold_decision_flips": { "END_to_HOLD": 5, "HOLD_to_END": 21, "count": 26, "false_interruptions_introduced": 4, "false_interruptions_resolved": 1, "missed_ends_introduced": 4, "missed_ends_resolved": 17, "rate": 0.07975460122699386, "unchanged_count": 300 } }, "trailing_silence_ms": 800 } ], "data_revision": "e564e2ac567f774d1880aa1db6ce97afb8c519b7", "data_scope": "preview-only: 3265 rows from train-00010-of-00083; official test remains sealed", "example_count": 326, "model_parameter_count": 151812, "model_window_seconds": 4.0, "negative_count": 166, "perturbation": { "conditions_ms": [ 0, 200, 400, 800 ], "policy": "Decode/resample the original suffix once, append valid zero-valued samples, then suffix-crop to the checkpoint window before the canonical frontend.", "sample_value": 0.0 }, "positive_count": 160, "privacy": { "aggregate_only": true, "contains_audio_or_transcripts": false, "contains_per_example_rows": false, "contains_raw_identifiers": false, "note": "The report contains aggregate paired statistics and file-level SHA-256 provenance only." }, "sample_rate": 16000, "schema_version": 1, "selection": { "max_examples": null, "num_workers": 0, "policy": "manifest-order prefix after split filtering", "shuffle": false }, "source_manifest": "data/processed/partial-iid-splits.jsonl", "source_manifest_sha256": "cae0765d18467186c0e890c35fd94c0fcc70ae3483b5a1ce31814dd8ffa37b50", "split": "validation", "threshold": 0.741000771522522, "training_status": "preview-only" }