{
  "text": "外放声音,离线出字幕。只用一个通用模型,实时字幕可能不够准,还得兼顾出字速度。先判断语言,再交给对应模型。边说边出临时字幕,停下才订稿。在Windows上,这两条命令分别抓电脑播放声,或把它和麦克风混合。再用这条命令开服务,就能在浏览器里看字幕。\n这是它的仪表盘,英文转写和字幕历史,都能直接看到。同批真实日语广播,其字符错误率3.8%,对比模型是13.8%。日语停说后,正式字幕典型约100毫秒出现,限制常驻模型时,内存不到2GB。近30天提交超100次。关注我,项目链接见评论区。",
  "language": "zh",
  "duration": 53.461375,
  "segments": [
    {
      "id": 0,
      "seek": 0,
      "start": 0.0,
      "end": 26.26,
      "text": "外放声音,离线出字幕。只用一个通用模型,实时字幕可能不够准,还得兼顾出字速度。先判断语言,再交给对应模型。边说边出临时字幕,停下才订稿。在Windows上,这两条命令分别抓电脑播放声,或把它和麦克风混合。再用这条命令开服务,就能在浏览器里看字幕。",
      "tokens": [
        50365,
        12022,
        12744,
        32045,
        18034,
        11,
        18614,
        119,
        16853,
        123,
        7781,
        22381,
        40211,
        1543,
        14003,
        9254,
        20182,
        19550,
        9254,
        41908,
        39823,
        11,
        24726,
        15729,
        22381,
        40211,
        16657,
        1960,
        1787,
        253,
        6336,
        228,
        11,
        14852,
        5916,
        2347,
        120,
        10178,
        122,
        7781,
        22381,
        31217,
        13127,
        1543,
        21176,
        1530,
        242,
        10108,
        2437,
        97,
        4307,
        255,
        5233,
        255,
        12009,
        11,
        8623,
        28455,
        23197,
        8713,
        44297,
        41908,
        39823,
        1543,
        40503,
        8090,
        40503,
        7781,
        940,
        112,
        15729,
        22381,
        40211,
        11,
        36135,
        4438,
        18888,
        7422,
        95,
        10415,
        123,
        1543,
        3581,
        45813,
        1509,
        5708,
        11,
        5562,
        36257,
        48837,
        25236,
        49061,
        6627,
        18453,
        43069,
        42182,
        27067,
        239,
        49993,
        12744,
        32045,
        11,
        19780,
        42061,
        12565,
        3994,
        99,
        24881,
        47209,
        48640,
        14245,
        1543,
        8623,
        9254,
        5562,
        48837,
        25236,
        49061,
        18937,
        27408,
        5087,
        94,
        11,
        3111,
        8225,
        3581,
        11038,
        237,
        6758,
        230,
        34386,
        15759,
        4200,
        22381,
        40211,
        1543,
        51678
      ],
      "temperature": 0,
      "avg_logprob": -0.057308588096563765,
      "compression_ratio": 1.2454873646209386,
      "no_speech_prob": 9.042999855879419e-12,
      "source": "whisper",
      "text_original": "外放声音,离线出字幕。只用一个通用模型,实时字幕可能不够准,还得兼顾出字速度。早尔先判断语言,再交给对应模型。边说边出临时字幕,停下才订稿。在Windows上,这两条命令分别抓电脑播放声,或把它和麦克风混合。再用这条命令开服务,就能在浏览器里看字幕。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 1,
      "seek": 2626,
      "start": 26.26,
      "end": 53.06,
      "text": "这是它的仪表盘,英文转写和字幕历史,都能直接看到。同批真实日语广播,其字符错误率3.8%,对比模型是13.8%。日语停说后,正式字幕典型约100毫秒出现,限制常驻模型时,内存不到2GB。近30天提交超100次。关注我,项目链接见评论区。",
      "tokens": [
        50365,
        27455,
        45224,
        1550,
        103,
        17571,
        5419,
        246,
        11,
        27869,
        17174,
        17819,
        105,
        5676,
        247,
        12565,
        22381,
        40211,
        5014,
        228,
        45399,
        11,
        7182,
        8225,
        43297,
        18032,
        1543,
        13089,
        3416,
        117,
        6303,
        24726,
        6890,
        5233,
        255,
        3509,
        123,
        49993,
        11,
        21176,
        1530,
        242,
        22381,
        5437,
        99,
        29900,
        5233,
        107,
        44866,
        18,
        13,
        23,
        8923,
        8713,
        11706,
        41908,
        39823,
        1541,
        7668,
        13,
        23,
        4,
        1543,
        6890,
        5233,
        255,
        36135,
        8090,
        13547,
        11,
        15789,
        27584,
        22381,
        40211,
        2347,
        116,
        39823,
        16853,
        99,
        6879,
        7256,
        104,
        20956,
        7781,
        20204,
        11,
        43446,
        25491,
        11279,
        26432,
        41908,
        39823,
        15729,
        11,
        34742,
        39781,
        35728,
        17,
        8769,
        1543,
        17463,
        3446,
        6135,
        20949,
        28455,
        19869,
        6879,
        9487,
        1543,
        28053,
        26432,
        1654,
        11,
        10178,
        117,
        11386,
        165,
        241,
        122,
        14468,
        23813,
        5233,
        226,
        7422,
        118,
        9937,
        118,
        1543,
        51705
      ],
      "temperature": 0,
      "avg_logprob": -0.05192097517160269,
      "compression_ratio": 1.162162162162162,
      "no_speech_prob": 1.3693358946742507e-11,
      "source": "whisper",
      "text_original": "这是它的仪表盘,英文转写和字幕历史,都能直接看到。同批真实日语广播,早尔字符错误率3.8%,对比模型是13.8%。日语停说后,正式字幕典型约100毫秒出现,限制常注模型时,内存不到2GB。近30天提交超100次。关注我,项目链接见评论区。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    }
  ],
  "transcript_selection": {
    "schema": "videonote.transcript_selection.v1",
    "status": "SELECTED",
    "requested_language": "zh",
    "duration": 53.461375,
    "reference_path": null,
    "reference_kind": null,
    "reference_language": null,
    "selected_source": "whisper_segmented",
    "decision_reason": "source_unavailable_or_untrusted_whisper_passed",
    "whisper_rejected": false,
    "source_profile": {
      "segment_count": 0,
      "duration_coverage": 0.0,
      "usable_full": false,
      "usable_partial": false,
      "hallucination_region_count": 0,
      "hallucination_regions": [],
      "language": {
        "requested": "zh",
        "matched": false
      }
    },
    "whisper_profile": {
      "segment_count": 2,
      "multilingual": false,
      "first_start": 0.0,
      "final_end": 53.06,
      "duration_coverage": 0.992492,
      "maximum_gap_seconds": 0.0,
      "inverted_count": 0,
      "language": {
        "requested": "zh",
        "matched": true,
        "latin": 9,
        "han": 190,
        "kana": 0,
        "hangul": 0
      },
      "hallucination_region_count": 0,
      "hallucination_regions": [],
      "usable_full": true,
      "usable_partial": true
    },
    "comparison": {
      "window_seconds": 30.0,
      "compared_window_count": 0,
      "median_score": 0.0,
      "divergent_window_count": 0,
      "divergent_ratio": 0.0,
      "divergent_windows": []
    },
    "timestamp_alignment": {
      "applied": false,
      "estimated_offset_seconds": 0.0,
      "matched_segment_count": 0,
      "match_scores": []
    },
    "caption_timing_repair": null,
    "filled_source_profile": null,
    "whisper_duration_guard": {
      "media_duration": 53.461,
      "rejected_segment_count": 0,
      "rejected_segments": []
    }
  },
  "presegmentation_correction": {
    "applied": true,
    "correction_count": 2,
    "deterministic_fix_count": 0
  }
}