{
  "text": "通常我給大家推薦一款AI工具的時候,往往我都習慣先描述它的痛點。但是這款微軟推出的AI工具就不一樣了。大家先看一下它的STAR數。53.5K啊,BRO。\n閉著眼就知道收藏了再說了。好,接著我們再看一個數據。能夠一次性處理長達60分鐘的音頻內容。什麼概念呢?你有用過其他ASR語音識別工具的話就知道了。\n很多都是處理不了長音頻,需要分割成短音頻才能繼續識別的。但是呢,這樣子一做的話,上下文就割裂了。這個和大語言模型的RAG的原理一樣啊。就是以前我們的大語言模型,它上下文不夠長嘛。它的Context Window不夠長。所以呢,我們需要先把一篇很長的文章啊,或者一個文檔啊,先叫它分割了。然後去做那個Word Embedding啊,丟給這個向量數據庫。然後等到用的時候呢。\n我們再去那個限量數據庫裡面去找,找到對應的內容。但是啊,因為你是一片一片,你分割了的。然後呢,你取回來的內容啊,可能已經缺失了很多上下文了。這個時候你再丟給這個大語言模型啊。往往你得到的不是你想要的結果。\n而這60分鐘的這個音頻呢,就相當於說你的大語言模型已經有很長的上下文了。你直接把一篇文檔丟給這個大語言模型,它就能全部識別到了。你不需要再去做分片。知道吧。\n這裡呢,我首先想到了一個應用場景啊,就是開會。比如說你一開一個小時的,直接將這一小時的這個錄音啊,丟給它總結。然後開完會了,就直接得到你的那個會議紀要了。\n又或者說你圍繞這個產品啊。做一個自己的一個硬件的產品也好。做一個自己的,比如錄音筆吧。做一個錄音筆。然後呢,你呢,開會的時候啊,直接錄音。然後呢,完了之後,直接呢,把這個啥?把這個會記要啊,直接發給你的各個同事。然後呢,發給他們的郵箱。是吧?這有很大的一個想像空間呢。看起來這個模型不大,大家可以看一下啊。這裡看到沒有?1.58G。然後呢,\nCPU就能跑了,GPU都不用買了。所以呢,隨便找個公司裡面的電腦啊,就能裝起來,就能玩啊。好,感興趣的話,趕緊玩起來吧。AI產品GG手,每天分享免費AI工具的技巧,關注!",
  "language": "zh",
  "duration": 144.54425,
  "segments": [
    {
      "id": 0,
      "seek": 0,
      "start": 0.0,
      "end": 14.6,
      "text": "通常我給大家推薦一款AI工具的時候,往往我都習慣先描述它的痛點。但是這款微軟推出的AI工具就不一樣了。大家先看一下它的STAR數。53.5K啊,BRO。",
      "tokens": [
        50365,
        19550,
        11279,
        1654,
        17798,
        6868,
        33597,
        27746,
        99,
        2257,
        48798,
        48698,
        23323,
        39806,
        20643,
        11,
        29510,
        29510,
        1654,
        7182,
        34025,
        12358,
        96,
        10108,
        11673,
        237,
        3316,
        108,
        45224,
        28737,
        8216,
        1543,
        11189,
        2664,
        48798,
        39152,
        14454,
        253,
        33597,
        7781,
        1546,
        48698,
        23323,
        39806,
        3111,
        1960,
        22977,
        2289,
        1543,
        6868,
        10108,
        28324,
        45224,
        6840,
        1899,
        30622,
        1543,
        19584,
        13,
        20,
        42,
        4905,
        11,
        33,
        7142,
        1543,
        51095
      ],
      "temperature": 0,
      "avg_logprob": -0.19896597685637296,
      "compression_ratio": 1.2643312101910829,
      "no_speech_prob": 9.252782567914508e-12,
      "source": "whisper"
    },
    {
      "id": 1,
      "seek": 0,
      "start": 14.6,
      "end": 29.6,
      "text": "閉著眼就知道收藏了再說了。好,接著我們再看一個數據。能夠一次性處理長達60分鐘的音頻內容。什麼概念呢?你有用過其他ASR語音識別工具的話就知道了。",
      "tokens": [
        51095,
        5196,
        231,
        19382,
        25281,
        3111,
        7758,
        18681,
        21782,
        237,
        2289,
        8623,
        4622,
        2289,
        1543,
        2131,
        11,
        14468,
        19382,
        5884,
        8623,
        4200,
        8990,
        30622,
        36841,
        1543,
        8225,
        31649,
        27505,
        21686,
        39289,
        13876,
        15353,
        33086,
        4550,
        32611,
        1546,
        18034,
        29940,
        28472,
        25750,
        1543,
        7598,
        26181,
        33679,
        6240,
        30,
        43320,
        9254,
        8816,
        9572,
        5000,
        3160,
        49,
        31348,
        18034,
        43143,
        16158,
        23323,
        39806,
        21358,
        3111,
        7758,
        2289,
        1543,
        51845
      ],
      "temperature": 0,
      "avg_logprob": -0.19896597685637296,
      "compression_ratio": 1.2643312101910829,
      "no_speech_prob": 9.252782567914508e-12,
      "source": "whisper"
    },
    {
      "id": 2,
      "seek": 2960,
      "start": 29.6,
      "end": 58.400000000000006,
      "text": "很多都是處理不了長音頻,需要分割成短音頻才能繼續識別的。但是呢,這樣子一做的話,上下文就割裂了。這個和大語言模型的RAG的原理一樣啊。就是以前我們的大語言模型,它上下文不夠長嘛。它的Context Window不夠長。所以呢,我們需要先把一篇很長的文章啊,或者一個文檔啊,先叫它分割了。然後去做那個Word Embedding啊,丟給這個向量數據庫。然後等到用的時候呢。",
      "tokens": [
        50365,
        20778,
        22796,
        39289,
        13876,
        47225,
        15353,
        18034,
        29940,
        11,
        35748,
        6627,
        5935,
        110,
        11336,
        5881,
        255,
        18034,
        29940,
        18888,
        8225,
        38459,
        43143,
        16158,
        1546,
        1543,
        11189,
        6240,
        11,
        24842,
        2257,
        10907,
        21358,
        11,
        4438,
        4438,
        17174,
        3111,
        5935,
        110,
        8083,
        224,
        2289,
        1543,
        6287,
        12565,
        3582,
        31348,
        12009,
        41908,
        39823,
        1546,
        3750,
        34,
        1546,
        19683,
        13876,
        22977,
        4905,
        1543,
        5620,
        44996,
        33641,
        3582,
        31348,
        12009,
        41908,
        39823,
        11,
        11284,
        5708,
        4438,
        17174,
        1960,
        31649,
        15353,
        20722,
        1543,
        45224,
        8547,
        394,
        301,
        44933,
        1960,
        31649,
        15353,
        1543,
        40685,
        11,
        5884,
        35748,
        10108,
        16075,
        2257,
        20878,
        229,
        4563,
        15353,
        1546,
        17174,
        11957,
        254,
        4905,
        11,
        31148,
        8990,
        17174,
        40741,
        242,
        4905,
        11,
        10108,
        19855,
        11284,
        6627,
        5935,
        110,
        2289,
        1543,
        10213,
        6734,
        10907,
        20754,
        54,
        765,
        24234,
        292,
        3584,
        4905,
        11,
        940,
        253,
        17798,
        6287,
        43446,
        26748,
        30622,
        36841,
        6346,
        104,
        1543,
        10213,
        10187,
        4511,
        9254,
        20643,
        6240,
        1543,
        51805
      ],
      "temperature": 0,
      "avg_logprob": -0.18821032947262392,
      "compression_ratio": 1.3324022346368716,
      "no_speech_prob": 1.3165710709139855e-11,
      "source": "whisper",
      "text_original": "很多都是處理不了長音頻,需要分割成短音頻才能繼續識別的。但是呢,這樣子一做的話,下下文就割裂了。這個和大語言模型的RAC的原理一樣啊。就是以前我們的大語言模型,它上下文不夠長嘛。它的Quantus Window不夠長。所以呢,我們需要先把一篇很長的文章啊,或者一個文檔啊,先叫它分割了。然後去做那個Word Embedding啊,丟給這個限量數據庫。然後等到用的時候呢。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 3,
      "seek": 5840,
      "start": 58.4,
      "end": 75.4,
      "text": "我們再去那個限量數據庫裡面去找,找到對應的內容。但是啊,因為你是一片一片,你分割了的。然後呢,你取回來的內容啊,可能已經缺失了很多上下文了。這個時候你再丟給這個大語言模型啊。往往你得到的不是你想要的結果。",
      "tokens": [
        50365,
        5884,
        8623,
        6734,
        20754,
        43446,
        26748,
        30622,
        36841,
        6346,
        104,
        32399,
        6734,
        25085,
        11,
        25085,
        4511,
        2855,
        20481,
        1546,
        28472,
        25750,
        1543,
        11189,
        4905,
        11,
        11471,
        32526,
        2257,
        16668,
        2257,
        16668,
        11,
        2166,
        6627,
        5935,
        110,
        2289,
        1546,
        1543,
        10213,
        6240,
        11,
        2166,
        29436,
        46556,
        1546,
        28472,
        25750,
        4905,
        11,
        16657,
        25338,
        38109,
        118,
        28564,
        2289,
        20778,
        5708,
        4438,
        17174,
        2289,
        1543,
        6287,
        14010,
        2166,
        8623,
        940,
        253,
        17798,
        6287,
        3582,
        31348,
        12009,
        41908,
        39823,
        4905,
        1543,
        29510,
        29510,
        2166,
        5916,
        4511,
        1546,
        7296,
        38386,
        4275,
        1546,
        35181,
        1543,
        51215
      ],
      "temperature": 0,
      "avg_logprob": -0.12223167829616095,
      "compression_ratio": 1.2595744680851064,
      "no_speech_prob": 2.026591915116338e-11,
      "source": "whisper"
    },
    {
      "id": 4,
      "seek": 7540,
      "start": 75.4,
      "end": 87.4,
      "text": "而這60分鐘的這個音頻呢,就相當於說你的大語言模型已經有很長的上下文了。你直接把一篇文檔丟給這個大語言模型,它就能全部識別到了。你不需要再去做分片。知道吧。",
      "tokens": [
        50365,
        11070,
        2664,
        4550,
        32611,
        1546,
        6287,
        18034,
        29940,
        6240,
        11,
        3111,
        15106,
        13118,
        19488,
        4622,
        18961,
        3582,
        31348,
        12009,
        41908,
        39823,
        25338,
        2412,
        4563,
        15353,
        1546,
        5708,
        4438,
        17174,
        2289,
        1543,
        2166,
        43297,
        16075,
        2257,
        20878,
        229,
        17174,
        40741,
        242,
        940,
        253,
        17798,
        6287,
        3582,
        31348,
        12009,
        41908,
        39823,
        11,
        11284,
        3111,
        8225,
        38714,
        43143,
        16158,
        21381,
        1543,
        2166,
        1960,
        35748,
        8623,
        6734,
        10907,
        6627,
        16668,
        1543,
        7758,
        6062,
        1543,
        50965
      ],
      "temperature": 0,
      "avg_logprob": -0.07836461067199707,
      "compression_ratio": 1.1957671957671958,
      "no_speech_prob": 2.1836105840167264e-11,
      "source": "whisper"
    },
    {
      "id": 5,
      "seek": 8740,
      "start": 87.4,
      "end": 99.4,
      "text": "這裡呢,我首先想到了一個應用場景啊,就是開會。比如說你一開一個小時的,直接將這一小時的這個錄音啊,丟給它總結。然後開完會了,就直接得到你的那個會議紀要了。",
      "tokens": [
        50365,
        18771,
        6240,
        11,
        1654,
        36490,
        7093,
        21381,
        8990,
        43772,
        235,
        14028,
        21446,
        50218,
        4905,
        11,
        5620,
        8949,
        6236,
        1543,
        36757,
        4622,
        2166,
        2257,
        8949,
        8990,
        46997,
        1546,
        11,
        43297,
        34168,
        32260,
        46997,
        1546,
        6287,
        9567,
        226,
        18034,
        4905,
        11,
        940,
        253,
        17798,
        11284,
        26575,
        17144,
        1543,
        10213,
        8949,
        14128,
        6236,
        2289,
        11,
        3111,
        43297,
        5916,
        4511,
        18961,
        20754,
        8949,
        6236,
        16958,
        4275,
        2289,
        1543,
        50965
      ],
      "temperature": 0,
      "avg_logprob": -0.17788544823141658,
      "compression_ratio": 1.2277777777777779,
      "no_speech_prob": 2.3383336933413545e-11,
      "source": "whisper",
      "text_original": "這裡呢,我首先想到了一個融入場景啊,就是開會。比如說你一開一個小時的,直接將這一小時的這個錄音啊,丟給它總結。然後開完會了,就直接得到你的那個開會記要了。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 6,
      "seek": 9940,
      "start": 99.4,
      "end": 129.32,
      "text": "又或者說你圍繞這個產品啊。做一個自己的一個硬件的產品也好。做一個自己的,比如錄音筆吧。做一個錄音筆。然後呢,你呢,開會的時候啊,直接錄音。然後呢,完了之後,直接呢,把這個啥?把這個會記要啊,直接發給你的各個同事。然後呢,發給他們的郵箱。是吧?這有很大的一個想像空間呢。看起來這個模型不大,大家可以看一下啊。這裡看到沒有?1.58G。然後呢,",
      "tokens": [
        50365,
        17047,
        31148,
        4622,
        2166,
        2523,
        235,
        23141,
        252,
        6287,
        33299,
        30246,
        4905,
        1543,
        10907,
        8990,
        17645,
        1546,
        8990,
        38114,
        105,
        20485,
        1546,
        33299,
        30246,
        6404,
        2131,
        1543,
        10907,
        8990,
        17645,
        1546,
        11,
        36757,
        9567,
        226,
        18034,
        7973,
        228,
        6062,
        1543,
        10907,
        8990,
        9567,
        226,
        18034,
        7973,
        228,
        1543,
        10213,
        6240,
        11,
        2166,
        6240,
        11,
        8949,
        6236,
        20643,
        4905,
        11,
        43297,
        9567,
        226,
        18034,
        1543,
        10213,
        6240,
        11,
        41665,
        26362,
        11,
        43297,
        6240,
        11,
        16075,
        6287,
        3284,
        98,
        30,
        16075,
        6287,
        6236,
        16958,
        4275,
        4905,
        11,
        43297,
        14637,
        17798,
        18961,
        17516,
        3338,
        13089,
        6973,
        1543,
        10213,
        6240,
        11,
        14637,
        17798,
        20486,
        1546,
        11178,
        113,
        11249,
        109,
        1543,
        1541,
        6062,
        30,
        2664,
        2412,
        4563,
        39156,
        8990,
        7093,
        12760,
        23322,
        11016,
        6240,
        1543,
        4200,
        21670,
        6287,
        41908,
        39823,
        1960,
        3582,
        11,
        6868,
        6723,
        28324,
        4905,
        1543,
        18771,
        18032,
        6963,
        30,
        16,
        13,
        20419,
        38,
        1543,
        10213,
        6240,
        11,
        51861
      ],
      "temperature": 0,
      "avg_logprob": -0.17717143833237206,
      "compression_ratio": 1.5526315789473684,
      "no_speech_prob": 2.2282269779294595e-11,
      "source": "whisper"
    },
    {
      "id": 7,
      "seek": 12932,
      "start": 129.32,
      "end": 141.32,
      "text": "CPU就能跑了,GPU都不用買了。所以呢,隨便找個公司裡面的電腦啊,就能裝起來,就能玩啊。好,感興趣的話,趕緊玩起來吧。AI產品GG手,每天分享免費AI工具的技巧,關注!",
      "tokens": [
        50365,
        34,
        8115,
        3111,
        8225,
        32585,
        2289,
        11,
        38,
        8115,
        7182,
        24384,
        31001,
        2289,
        1543,
        40685,
        11,
        48890,
        27364,
        25085,
        3338,
        13545,
        32981,
        32399,
        1546,
        20545,
        21184,
        99,
        4905,
        11,
        3111,
        8225,
        44453,
        21670,
        11,
        3111,
        8225,
        19912,
        4905,
        1543,
        2131,
        11,
        9709,
        44089,
        39835,
        21358,
        11,
        9759,
        243,
        30172,
        19912,
        21670,
        6062,
        1543,
        48698,
        33299,
        30246,
        27561,
        11389,
        11,
        23664,
        6135,
        30855,
        2347,
        235,
        50071,
        48698,
        23323,
        39806,
        1546,
        32502,
        43605,
        11,
        14899,
        26432,
        0,
        50965
      ],
      "temperature": 0,
      "avg_logprob": -0.16701302161583534,
      "compression_ratio": 1.097938144329897,
      "no_speech_prob": 2.364201369398078e-11,
      "source": "whisper"
    }
  ],
  "transcript_selection": {
    "schema": "videonote.transcript_selection.v1",
    "status": "SELECTED",
    "requested_language": "zh",
    "duration": 144.54425,
    "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": 8,
      "multilingual": false,
      "first_start": 0.0,
      "final_end": 141.32,
      "duration_coverage": 0.977694,
      "maximum_gap_seconds": 0.0,
      "inverted_count": 0,
      "language": {
        "requested": "zh",
        "matched": true,
        "latin": 57,
        "han": 687,
        "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": 144.544,
      "rejected_segment_count": 0,
      "rejected_segments": []
    }
  },
  "presegmentation_correction": {
    "applied": true,
    "correction_count": 2,
    "deterministic_fix_count": 0
  }
}