{
  "text": "哎,會議裡幾個人同時說話,普通語音識別的問題不只是聽不清,更是不知道該聽誰。如果你只需要其中一個人的文字,先把每個人都分離出來,未必是最直接的路。\nXiaomi Cocktail ASR-1給出了另一個思路。先輸入一段目標人的參考聲音,再輸入單人或多人混合語音。參考聲音就像一把聲文鑰匙,模型直接只轉錄對應的說話人。\n機器說清了，先聽一組官方演示。第一段是目標人的參考聲音，第二段是待識別的混合語音。\n我想註冊一下聲文,這是一條註冊音頻。\n人們設計AI配方的過程,本質上主要還是一個不斷試錯的過程。\n模型最後輸出的是,人們設計AI配方的過程,本質上主要還是一個不斷試錯的過程。\n這裡不是先分出每一路音頻,而是用參考聲音直接約束轉寫目標。\n聽到效果後，再看真實會議數據。官方報告中，這個模型在AM&SDM上的詞錯率是21.81%，在EliminatingFar上是20.63%，對比表裡都明顯低於直接處理混合語音的通用模型。\n但只會轉路還不夠,如果參考聲音的人根本沒在待識別音頻裡,模型應該保持沉默。\n下面再聽一組副樣本。\n嗨同學。\n嗨有關系。\n這組副樣本的官方結果是空。\n三個正樣本測試集上，句實率分別是79.59%，75.35%和68.54%，但句實不是100%，正樣本也有最高0.73%的物據率。\n選型時要同時看詞錯率、句實率和誤識率。\n開始試跑,只需要四個依賴,兩段音頻和一次模型調用。\n準備一段目標人參考音頻和一段待識別音頻。\n輸入期望是16000赫茲單聲道，成功後應該只返回目標人文本。\n再換一段不相關的參考聲音,預期結果是空。\n最後的選型結論很清楚，指定受訪者轉錄，以登記用戶語音入口或多人環境跟蹤固定說話人。\n這很對提。\n需要保留所有發言人的完整會議,就應該用說話人日誌或分離流程。\n官方示例使用約80億參數模型、CUDA和遠程自定義代碼,先做小樣本驗證再接入。\n這裡是智聲工坊。\n我們在智聲工坊。\n我們在智聲工坊。\n我們在智聲工坊。",
  "language": "zh",
  "duration": 131.690688,
  "segments": [
    {
      "id": 0,
      "seek": 0,
      "start": 0.0,
      "end": 10.0,
      "text": "哎,會議裡幾個人同時說話,普通語音識別的問題不只是聽不清,更是不知道該聽誰。如果你只需要其中一個人的文字,先把每個人都分離出來,未必是最直接的路。",
      "tokens": [
        50365,
        26755,
        11,
        6236,
        24686,
        11066,
        23575,
        40104,
        13089,
        6611,
        4622,
        11103,
        11,
        29993,
        19550,
        31348,
        18034,
        43143,
        16158,
        1546,
        17197,
        1960,
        36859,
        21524,
        1960,
        21784,
        11,
        19002,
        1541,
        17572,
        22033,
        21524,
        20757,
        1543,
        45669,
        14003,
        35748,
        9572,
        5975,
        8990,
        4035,
        1546,
        17174,
        22381,
        11,
        10108,
        16075,
        23664,
        40104,
        7182,
        6627,
        32662,
        29741,
        11,
        29954,
        28531,
        1541,
        8661,
        43297,
        1546,
        24658,
        1543,
        50865
      ],
      "temperature": 0,
      "avg_logprob": -0.11201262343776681,
      "compression_ratio": 1.4164456233421752,
      "no_speech_prob": 9.735798105203752e-12,
      "source": "whisper"
    },
    {
      "id": 1,
      "seek": 0,
      "start": 10.0,
      "end": 22.0,
      "text": "Xiaomi Cocktail ASR-1給出了另一個思路。先輸入一段目標人的參考聲音,再輸入單人或多人混合語音。參考聲音就像一把聲文鑰匙,模型直接只轉錄對應的說話人。",
      "tokens": [
        50865,
        55,
        654,
        9220,
        39410,
        14430,
        7469,
        49,
        12,
        16,
        17798,
        7781,
        2289,
        22762,
        8990,
        8870,
        24658,
        1543,
        10108,
        12290,
        116,
        14028,
        2257,
        28427,
        11386,
        46299,
        4035,
        1546,
        47375,
        26504,
        29454,
        18034,
        11,
        8623,
        12290,
        116,
        14028,
        24227,
        4035,
        19780,
        6392,
        4035,
        48640,
        14245,
        31348,
        18034,
        1543,
        47375,
        26504,
        29454,
        18034,
        3111,
        12760,
        2257,
        16075,
        29454,
        17174,
        165,
        239,
        108,
        9937,
        247,
        11,
        41908,
        39823,
        43297,
        14003,
        32739,
        9567,
        226,
        2855,
        20481,
        1546,
        4622,
        11103,
        4035,
        1543,
        51465
      ],
      "temperature": 0,
      "avg_logprob": -0.11201262343776681,
      "compression_ratio": 1.4164456233421752,
      "no_speech_prob": 9.735798105203752e-12,
      "source": "whisper"
    },
    {
      "id": 2,
      "seek": 0,
      "start": 22.0,
      "end": 29.0,
      "text": "機器說清了，先聽一組官方演示。第一段是目標人的參考聲音，第二段是待識別的混合語音。",
      "tokens": [
        51465,
        17543,
        7626,
        4622,
        21784,
        2289,
        11,
        10108,
        21524,
        2257,
        30052,
        31929,
        9249,
        31382,
        25696,
        1543,
        18049,
        28427,
        1541,
        11386,
        46299,
        4035,
        1546,
        47375,
        26504,
        29454,
        18034,
        11,
        19693,
        28427,
        1541,
        32379,
        43143,
        16158,
        1546,
        48640,
        14245,
        31348,
        18034,
        1543,
        51815
      ],
      "temperature": 0,
      "avg_logprob": -0.11201262343776681,
      "compression_ratio": 1.4164456233421752,
      "no_speech_prob": 9.735798105203752e-12,
      "source": "whisper",
      "text_original": "機子說清了,先聽一組官方演示。第一段是目標人的參考聲音,第二段是帶識別的混合語音。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 3,
      "seek": 3000,
      "start": 30.0,
      "end": 35.0,
      "text": "我想註冊一下聲文,這是一條註冊音頻。",
      "tokens": [
        50365,
        25246,
        5396,
        119,
        5676,
        232,
        8861,
        29454,
        17174,
        11,
        16600,
        2257,
        31106,
        5396,
        119,
        5676,
        232,
        18034,
        29940,
        1543,
        50615
      ],
      "temperature": 0,
      "avg_logprob": -0.06891813116558528,
      "compression_ratio": 1.588235294117647,
      "no_speech_prob": 1.885463486728245e-11,
      "source": "whisper"
    },
    {
      "id": 4,
      "seek": 3000,
      "start": 35.0,
      "end": 41.0,
      "text": "人們設計AI配方的過程,本質上主要還是一個不斷試錯的過程。",
      "tokens": [
        50615,
        4035,
        4623,
        39035,
        30114,
        48698,
        38846,
        9249,
        1546,
        8816,
        29649,
        11,
        8802,
        43450,
        5708,
        13557,
        4275,
        25583,
        8990,
        1960,
        4307,
        115,
        22099,
        13133,
        1546,
        8816,
        29649,
        1543,
        50915
      ],
      "temperature": 0,
      "avg_logprob": -0.06891813116558528,
      "compression_ratio": 1.588235294117647,
      "no_speech_prob": 1.885463486728245e-11,
      "source": "whisper"
    },
    {
      "id": 5,
      "seek": 3000,
      "start": 41.0,
      "end": 48.0,
      "text": "模型最後輸出的是,人們設計AI配方的過程,本質上主要還是一個不斷試錯的過程。",
      "tokens": [
        50915,
        41908,
        39823,
        20578,
        12290,
        116,
        7781,
        24620,
        11,
        4035,
        4623,
        39035,
        30114,
        48698,
        38846,
        9249,
        1546,
        8816,
        29649,
        11,
        8802,
        43450,
        5708,
        13557,
        4275,
        25583,
        8990,
        1960,
        4307,
        115,
        22099,
        13133,
        1546,
        8816,
        29649,
        1543,
        51265
      ],
      "temperature": 0,
      "avg_logprob": -0.06891813116558528,
      "compression_ratio": 1.588235294117647,
      "no_speech_prob": 1.885463486728245e-11,
      "source": "whisper"
    },
    {
      "id": 6,
      "seek": 3000,
      "start": 48.0,
      "end": 53.0,
      "text": "這裡不是先分出每一路音頻,而是用參考聲音直接約束轉寫目標。",
      "tokens": [
        51265,
        18771,
        7296,
        10108,
        6627,
        7781,
        23664,
        2257,
        24658,
        18034,
        29940,
        11,
        11070,
        1541,
        9254,
        47375,
        26504,
        29454,
        18034,
        43297,
        30889,
        44320,
        32739,
        4510,
        104,
        11386,
        46299,
        1543,
        51515
      ],
      "temperature": 0,
      "avg_logprob": -0.06891813116558528,
      "compression_ratio": 1.588235294117647,
      "no_speech_prob": 1.885463486728245e-11,
      "source": "whisper"
    },
    {
      "id": 7,
      "seek": 5300,
      "start": 53.0,
      "end": 67.0,
      "text": "聽到效果後，再看真實會議數據。官方報告中，這個模型在AM&SDM上的詞錯率是21.81%，在EliminatingFar上是20.63%，對比表裡都明顯低於直接處理混合語音的通用模型。",
      "tokens": [
        50365,
        21524,
        4511,
        43076,
        9319,
        5661,
        11,
        8623,
        4200,
        6303,
        10376,
        6236,
        24686,
        30622,
        36841,
        1543,
        31929,
        9249,
        17803,
        16846,
        5975,
        11,
        6287,
        41908,
        39823,
        3581,
        2865,
        5,
        23969,
        44,
        5708,
        1546,
        6716,
        252,
        13133,
        44866,
        1541,
        4436,
        13,
        32875,
        8923,
        3581,
        36,
        4197,
        1748,
        37,
        289,
        5708,
        1541,
        2009,
        13,
        29491,
        8923,
        2855,
        11706,
        17571,
        11066,
        7182,
        11100,
        46605,
        41377,
        19488,
        43297,
        39289,
        13876,
        48640,
        14245,
        31348,
        18034,
        1546,
        19550,
        9254,
        41908,
        39823,
        1543,
        51065
      ],
      "temperature": 0,
      "avg_logprob": -0.16579674451779097,
      "compression_ratio": 0.9950738916256158,
      "no_speech_prob": 2.0118370511190697e-11,
      "source": "whisper",
      "text_original": "聽到效果後,再看真實會議數據。官方報告中,這個模型在AM&SDM上的詞錯率是21.81%,在ElimitingFar上是20.63%,對比表裡都明顯低於直接處理混合語音的通用模型。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 8,
      "seek": 6700,
      "start": 67.0,
      "end": 74.0,
      "text": "但只會轉路還不夠,如果參考聲音的人根本沒在待識別音頻裡,模型應該保持沉默。",
      "tokens": [
        50365,
        8395,
        14003,
        6236,
        32739,
        24658,
        7824,
        1960,
        31649,
        11,
        13119,
        47375,
        26504,
        29454,
        18034,
        29979,
        31337,
        8802,
        4913,
        3581,
        18390,
        43143,
        16158,
        18034,
        29940,
        11066,
        11,
        41908,
        39823,
        26087,
        24302,
        17694,
        3308,
        231,
        6173,
        246,
        1543,
        50715
      ],
      "temperature": 0,
      "avg_logprob": -0.15461736525932368,
      "compression_ratio": 1.2153284671532847,
      "no_speech_prob": 2.001824921105122e-11,
      "source": "whisper"
    },
    {
      "id": 9,
      "seek": 6700,
      "start": 74.0,
      "end": 76.0,
      "text": "下面再聽一組副樣本。",
      "tokens": [
        50715,
        47150,
        8623,
        21524,
        2257,
        30052,
        5935,
        107,
        6180,
        8802,
        1543,
        50815
      ],
      "temperature": 0,
      "avg_logprob": -0.15461736525932368,
      "compression_ratio": 1.2153284671532847,
      "no_speech_prob": 2.001824921105122e-11,
      "source": "whisper"
    },
    {
      "id": 10,
      "seek": 6700,
      "start": 76.0,
      "end": 77.0,
      "text": "嗨同學。",
      "tokens": [
        50815,
        6083,
        101,
        13089,
        21372,
        1543,
        50865
      ],
      "temperature": 0,
      "avg_logprob": -0.15461736525932368,
      "compression_ratio": 1.2153284671532847,
      "no_speech_prob": 2.001824921105122e-11,
      "source": "whisper"
    },
    {
      "id": 11,
      "seek": 6700,
      "start": 77.0,
      "end": 80.0,
      "text": "嗨有關系。",
      "tokens": [
        50865,
        6083,
        101,
        2412,
        14899,
        25368,
        1543,
        51015
      ],
      "temperature": 0,
      "avg_logprob": -0.15461736525932368,
      "compression_ratio": 1.2153284671532847,
      "no_speech_prob": 2.001824921105122e-11,
      "source": "whisper"
    },
    {
      "id": 12,
      "seek": 6700,
      "start": 80.0,
      "end": 82.0,
      "text": "這組副樣本的官方結果是空。",
      "tokens": [
        51015,
        2664,
        30052,
        5935,
        107,
        6180,
        8802,
        1546,
        31929,
        9249,
        35181,
        1541,
        23322,
        1543,
        51115
      ],
      "temperature": 0,
      "avg_logprob": -0.15461736525932368,
      "compression_ratio": 1.2153284671532847,
      "no_speech_prob": 2.001824921105122e-11,
      "source": "whisper"
    },
    {
      "id": 13,
      "seek": 6700,
      "start": 82.0,
      "end": 93.0,
      "text": "三個正樣本測試集上，句實率分別是79.59%，75.35%和68.54%，但句實不是100%，正樣本也有最高0.73%的物據率。",
      "tokens": [
        51115,
        10960,
        3338,
        5935,
        107,
        6180,
        8802,
        9592,
        105,
        22099,
        26020,
        5708,
        11,
        34592,
        10376,
        44866,
        6627,
        16158,
        1541,
        32042,
        13,
        19600,
        8923,
        11901,
        13,
        8794,
        4,
        12565,
        27102,
        13,
        19563,
        8923,
        8395,
        34592,
        10376,
        7296,
        6879,
        8923,
        15789,
        6180,
        8802,
        6404,
        2412,
        8661,
        12979,
        15,
        13,
        33396,
        4,
        1546,
        23516,
        36841,
        44866,
        1543,
        51665
      ],
      "temperature": 0,
      "avg_logprob": -0.15461736525932368,
      "compression_ratio": 1.2153284671532847,
      "no_speech_prob": 2.001824921105122e-11,
      "source": "whisper",
      "text_original": "三個副樣本測試集上,句實率分別是79.59%,75.35%和68.54%,但句實不是100%,正樣本也有最高0.73%的物據率。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 14,
      "seek": 9300,
      "start": 93.0,
      "end": 97.0,
      "text": "選型時要同時看詞錯率、句實率和誤識率。",
      "tokens": [
        50365,
        21625,
        8082,
        6611,
        4275,
        13089,
        6611,
        4200,
        6716,
        252,
        13133,
        44866,
        1231,
        34592,
        10376,
        44866,
        12565,
        23516,
        36841,
        44866,
        1543,
        50565
      ],
      "temperature": 0,
      "avg_logprob": -0.09947648094695749,
      "compression_ratio": 1.364321608040201,
      "no_speech_prob": 1.7557613349317158e-11,
      "source": "whisper",
      "text_original": "選行時要同時看詞錯率、句實率和物據率。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 15,
      "seek": 9300,
      "start": 97.0,
      "end": 100.0,
      "text": "開始試跑,只需要四個依賴,兩段音頻和一次模型調用。",
      "tokens": [
        50565,
        21017,
        22099,
        32585,
        11,
        14003,
        35748,
        19425,
        3338,
        3254,
        251,
        164,
        7062,
        11,
        16313,
        28427,
        18034,
        29940,
        12565,
        27505,
        41908,
        39823,
        32943,
        9254,
        1543,
        50715
      ],
      "temperature": 0,
      "avg_logprob": -0.09947648094695749,
      "compression_ratio": 1.364321608040201,
      "no_speech_prob": 1.7557613349317158e-11,
      "source": "whisper"
    },
    {
      "id": 16,
      "seek": 9300,
      "start": 100.0,
      "end": 103.0,
      "text": "準備一段目標人參考音頻和一段待識別音頻。",
      "tokens": [
        50715,
        36241,
        2257,
        28427,
        11386,
        46299,
        4035,
        47375,
        26504,
        18034,
        29940,
        12565,
        2257,
        28427,
        19105,
        43143,
        16158,
        18034,
        29940,
        1543,
        50865
      ],
      "temperature": 0,
      "avg_logprob": -0.09947648094695749,
      "compression_ratio": 1.364321608040201,
      "no_speech_prob": 1.7557613349317158e-11,
      "source": "whisper",
      "text_original": "準備一段目標人參考音頻和一段代識別音頻。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 17,
      "seek": 9300,
      "start": 103.0,
      "end": 107.0,
      "text": "輸入期望是16000赫茲單聲道，成功後應該只返回目標人文本。",
      "tokens": [
        50865,
        12290,
        116,
        14028,
        16786,
        22694,
        1541,
        6866,
        1360,
        5266,
        104,
        2347,
        117,
        29454,
        6025,
        11,
        11336,
        32311,
        5661,
        26087,
        14003,
        3316,
        242,
        8350,
        11386,
        46299,
        4035,
        17174,
        8802,
        1543,
        51065
      ],
      "temperature": 0,
      "avg_logprob": -0.09947648094695749,
      "compression_ratio": 1.364321608040201,
      "no_speech_prob": 1.7557613349317158e-11,
      "source": "whisper",
      "text_original": "輸入期望是16000赫兹聲道,成功後應該只返回目標人文本。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 18,
      "seek": 9300,
      "start": 107.0,
      "end": 110.0,
      "text": "再換一段不相關的參考聲音,預期結果是空。",
      "tokens": [
        51065,
        8623,
        36338,
        2257,
        28427,
        1960,
        15106,
        14899,
        1546,
        47375,
        26504,
        29454,
        18034,
        11,
        30156,
        16786,
        35181,
        1541,
        23322,
        1543,
        51215
      ],
      "temperature": 0,
      "avg_logprob": -0.09947648094695749,
      "compression_ratio": 1.364321608040201,
      "no_speech_prob": 1.7557613349317158e-11,
      "source": "whisper"
    },
    {
      "id": 19,
      "seek": 9300,
      "start": 110.0,
      "end": 116.0,
      "text": "最後的選型結論很清楚，指定受訪者轉錄，以登記用戶語音入口或多人環境跟蹤固定說話人。",
      "tokens": [
        51215,
        20578,
        1546,
        21625,
        8082,
        17144,
        29340,
        4563,
        38433,
        11,
        25922,
        12088,
        23151,
        5396,
        103,
        12444,
        32739,
        24658,
        11,
        3588,
        25874,
        16958,
        9254,
        1486,
        114,
        31348,
        18034,
        14028,
        18144,
        19780,
        6392,
        4035,
        40063,
        108,
        47885,
        9678,
        41166,
        103,
        3919,
        118,
        12088,
        4622,
        11103,
        4035,
        1543,
        51515
      ],
      "temperature": 0,
      "avg_logprob": -0.09947648094695749,
      "compression_ratio": 1.364321608040201,
      "no_speech_prob": 1.7557613349317158e-11,
      "source": "whisper",
      "text_original": "最後的選行結論很清楚,指定受訪者轉路,以登記用戶語音入口或多人環境跟踪固定說話人。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 20,
      "seek": 9300,
      "start": 116.0,
      "end": 117.0,
      "text": "這很對提。",
      "tokens": [
        51515,
        5000,
        4563,
        2855,
        20949,
        1543,
        51565
      ],
      "temperature": 0,
      "avg_logprob": -0.09947648094695749,
      "compression_ratio": 1.364321608040201,
      "no_speech_prob": 1.7557613349317158e-11,
      "source": "whisper",
      "text_original": "他很對提。",
      "correction_applied": true,
      "correction_source": "llm_correct_a_presegmentation"
    },
    {
      "id": 21,
      "seek": 9300,
      "start": 117.0,
      "end": 121.0,
      "text": "需要保留所有發言人的完整會議,就應該用說話人日誌或分離流程。",
      "tokens": [
        51565,
        35748,
        24302,
        24456,
        39300,
        14637,
        12009,
        4035,
        1546,
        14128,
        27662,
        6236,
        24686,
        11,
        3111,
        26087,
        9254,
        4622,
        11103,
        4035,
        6890,
        3549,
        234,
        19780,
        6627,
        32662,
        27854,
        29649,
        1543,
        51765
      ],
      "temperature": 0,
      "avg_logprob": -0.09947648094695749,
      "compression_ratio": 1.364321608040201,
      "no_speech_prob": 1.7557613349317158e-11,
      "source": "whisper"
    },
    {
      "id": 22,
      "seek": 12100,
      "start": 121.0,
      "end": 127.0,
      "text": "官方示例使用約80億參數模型、CUDA和遠程自定義代碼,先做小樣本驗證再接入。",
      "tokens": [
        50365,
        31929,
        9249,
        25696,
        17797,
        22982,
        9254,
        30889,
        4702,
        25459,
        47375,
        30622,
        41908,
        39823,
        1231,
        25864,
        7509,
        12565,
        34489,
        29649,
        9722,
        12088,
        37430,
        19105,
        16337,
        120,
        11,
        10108,
        10907,
        7322,
        6180,
        8802,
        24023,
        245,
        11823,
        231,
        8623,
        14468,
        14028,
        1543,
        50665
      ],
      "temperature": 0,
      "avg_logprob": -0.17029684561270256,
      "compression_ratio": 0.9202898550724637,
      "no_speech_prob": 2.0532789010707653e-11,
      "source": "whisper"
    },
    {
      "id": 23,
      "seek": 12100,
      "start": 127.0,
      "end": 128.0,
      "text": "這裡是智聲工坊。",
      "tokens": [
        50665,
        18771,
        1541,
        5094,
        118,
        29454,
        23323,
        14872,
        232,
        1543,
        50715
      ],
      "temperature": 0,
      "avg_logprob": -0.17029684561270256,
      "compression_ratio": 0.9202898550724637,
      "no_speech_prob": 2.0532789010707653e-11,
      "source": "whisper"
    },
    {
      "id": 24,
      "seek": 12800,
      "start": 128.0,
      "end": 130.0,
      "text": "我們在智聲工坊。",
      "tokens": [
        50365,
        5884,
        3581,
        5094,
        118,
        29454,
        23323,
        14872,
        232,
        1543,
        50465
      ],
      "temperature": 0,
      "avg_logprob": -0.20211913446674432,
      "compression_ratio": 12.15,
      "no_speech_prob": 6.076185388170785e-11,
      "source": "whisper"
    },
    {
      "id": 25,
      "seek": 12800,
      "start": 130.0,
      "end": 131.0,
      "text": "我們在智聲工坊。",
      "tokens": [
        50465,
        5884,
        3581,
        5094,
        118,
        29454,
        23323,
        14872,
        232,
        1543,
        50515
      ],
      "temperature": 0,
      "avg_logprob": -0.20211913446674432,
      "compression_ratio": 12.15,
      "no_speech_prob": 6.076185388170785e-11,
      "source": "whisper"
    },
    {
      "id": 26,
      "seek": 12800,
      "start": 131.0,
      "end": 131.690688,
      "text": "我們在智聲工坊。",
      "tokens": [
        50515,
        5884,
        3581,
        5094,
        118,
        29454,
        23323,
        14872,
        232,
        1543,
        50565
      ],
      "temperature": 0,
      "avg_logprob": -0.20211913446674432,
      "compression_ratio": 12.15,
      "no_speech_prob": 6.076185388170785e-11,
      "source": "whisper"
    }
  ],
  "transcript_selection": {
    "schema": "videonote.transcript_selection.v1",
    "status": "SELECTED",
    "requested_language": "zh",
    "duration": 131.690688,
    "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": 27,
      "multilingual": false,
      "first_start": 0.0,
      "final_end": 131.691,
      "duration_coverage": 1.0,
      "maximum_gap_seconds": 1.0,
      "inverted_count": 0,
      "language": {
        "requested": "zh",
        "matched": true,
        "latin": 42,
        "han": 636,
        "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": 131.691,
      "rejected_segment_count": 17,
      "rejected_segments": [
        {
          "index": 27,
          "start": 132.0,
          "end": 133.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 28,
          "start": 133.0,
          "end": 134.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 29,
          "start": 134.0,
          "end": 135.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 30,
          "start": 135.0,
          "end": 136.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 31,
          "start": 136.0,
          "end": 137.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 32,
          "start": 137.0,
          "end": 138.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 33,
          "start": 138.0,
          "end": 139.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 34,
          "start": 139.0,
          "end": 140.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 35,
          "start": 140.0,
          "end": 141.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 36,
          "start": 141.0,
          "end": 142.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 37,
          "start": 142.0,
          "end": 143.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 38,
          "start": 143.0,
          "end": 144.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 39,
          "start": 144.0,
          "end": 145.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 40,
          "start": 145.0,
          "end": 146.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 41,
          "start": 146.0,
          "end": 147.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 42,
          "start": 147.0,
          "end": 148.0,
          "reason": "starts_at_or_after_media_end"
        },
        {
          "index": 43,
          "start": 148.0,
          "end": 149.0,
          "reason": "starts_at_or_after_media_end"
        }
      ]
    }
  },
  "presegmentation_correction": {
    "applied": true,
    "correction_count": 8,
    "deterministic_fix_count": 0
  }
}