{"text": "利用Whisperger给日语电影配中文字幕的终极教程,如何让人工校正最轻松?1.下载Whisperger模式选择双通道模式。PAS1选择Fidelity模式Balance灵敏度加OutTalk能量检测加10.5分段更细,不会字幕过长。PAS2选择Balance模式Aggressive灵敏度加Colero加Colero6.2V。第一遍保证模式才是准确度的大头,准确度占比80%,第二遍只是填充的,以弥补字幕不全。所以PAS1保证模式必选精细度最高的OutTalk加10。而PAS2平衡模式准确度更低,仅仅为了填充,所以选精细度不高的Colero,保证模式更强。而且比平衡模式准确度强太多,保证模式是FP16,平衡模式是影的8。所以必选以PAS1为主,PAS2填充,这样准确度更高后期人工校正更轻松。2. 模型必选Large VR而不是Large V3。VR等于日语专精型,V3等于全球平均型,转入日语首选V2。V3做了两个大升级,梅尔平谱从80-128,日语更容易收噪音影响。词调Token扩容加了很多小语种,日语全重下降。这两个改动对日语全是浮悠化。V3更改编,日语最受不了幻觉。V3为了流畅度更容易脑补内容,日语一脑补就是语法错,禁语错,注词。错VR更保守,不乱编日语反而更稳。3. Whisper J A V尽管能把Large VR的幻觉降低,但是实测幻觉依然存在。可以使用AI翻译成中文后,用Subtitle Edit软件打开该字幕。进入多重替换,用正则表达式或者正常字符替换添加幻觉词消除。Large VR转写加AI翻译后的幻觉并非随机乱码,而是有明确的高频固定模式。多余的冗于双引号,莫名出现的人物前缀,固定同音措辞等。色Title Edit的多重替换,支持一次性导入整份规则库,也就是我提到的幻觉词表。双引号修正规则,全局批量执行,还可以反复运行二到三次彻底清理。全程不需要逐行手动修改,效率远高于纯人工校对。Large VR的幻觉本质是训练数据高度依赖YouTube预料。模型在静音、噪音、低质性读音频段会自动脑补出训练里最高频的YouTube互动画书,属于模型先天特性,并非操作故障。4. 关于人工修正,经过色Pyto Edit清洗以上步骤后,修改起来还是比较轻松。我愿称之为目前人工修正最轻松,幻觉最少的字幕软件。一般300行5到10分钟左右,600行10到15分钟左右,快一点的话就5分钟,看你是不是想认真改。关于笔记本设备,推荐3070TI笔记本,拥有5888个流处理器,是性价比最高的AI设备选择。转录速度比4060亿4070万5060笔记本快,而且价格也不贵。只要把温度控制好,就能最大程度发挥效率,节省时间。5.字幕的质量还算可以,精确度非常高。当然尽管用了双通道模式避免漏掉字幕,但或多或少还是会漏掉一点字幕,不过瑕不掩瑜。如果强行上保真模式最高灵敏度,首先幻觉会直线上升。其次是字幕人工修正更辛苦,两者直接影响后面的准确度,而且时间太漫长得不偿失。为了这一点漏掉的字幕而放弃最终的准确度,没必要。假如是一小时时评600行,Whisper这保真模式转录需要一小时20-30分钟左右,3070TI笔记本。AI翻译也就2-3分钟,多重替换也就10几秒。人工校正一般10-15分钟。整体上看最耗费时间的还是转录,所以可以把转录时间留给凌晨。一般凌晨0点转录,到早上8点,一般就能转录好5个视频。视频时长都是一小时左右,不影响白天玩电脑。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 6.0, "text": "利用Whisperger给日语电影配中文字幕的终极教程,如何让人工校正最轻松?", "tokens": [50365, 23700, 9254, 2471, 271, 610, 1321, 23197, 6890, 5233, 255, 42182, 16820, 38846, 5975, 17174, 22381, 40211, 1546, 10115, 230, 7360, 223, 21936, 29649, 11, 43526, 33650, 4035, 23323, 49684, 15789, 8661, 17819, 119, 36897, 30, 50665], "temperature": 0, "avg_logprob": -0.17544205622239548, "compression_ratio": 1.2090032154340835, "no_speech_prob": 9.655437040179127e-12}, {"id": 1, "seek": 0, "start": 6.0, "end": 10.0, "text": "1.下载Whisperger模式选择双通道模式。", "tokens": [50665, 16, 13, 4438, 17819, 121, 2471, 271, 610, 1321, 41908, 27584, 2215, 231, 6852, 102, 2129, 234, 19550, 6025, 41908, 27584, 1543, 50865], "temperature": 0, "avg_logprob": -0.17544205622239548, "compression_ratio": 1.2090032154340835, "no_speech_prob": 9.655437040179127e-12}, {"id": 2, "seek": 0, "start": 10.0, "end": 17.0, "text": "PAS1选择Fidelity模式Balance灵敏度加OutTalk能量检测加10.5分段更细,不会字幕过长。", "tokens": [50865, 47, 3160, 16, 2215, 231, 6852, 102, 37, 33343, 41908, 27584, 33, 304, 719, 13290, 113, 7017, 237, 13127, 9990, 28353, 33210, 8225, 26748, 16407, 222, 11038, 233, 9990, 3279, 13, 20, 6627, 28427, 19002, 10115, 228, 11, 47928, 22381, 40211, 16866, 32271, 1543, 51215], "temperature": 0, "avg_logprob": -0.17544205622239548, "compression_ratio": 1.2090032154340835, "no_speech_prob": 9.655437040179127e-12}, {"id": 3, "seek": 0, "start": 17.0, "end": 22.0, "text": "PAS2选择Balance模式Aggressive灵敏度加Colero加Colero6.2V。", "tokens": [51215, 47, 3160, 17, 2215, 231, 6852, 102, 33, 304, 719, 41908, 27584, 32, 1615, 22733, 13290, 113, 7017, 237, 13127, 9990, 34, 401, 2032, 9990, 34, 401, 2032, 21, 13, 17, 53, 1543, 51465], "temperature": 0, "avg_logprob": -0.17544205622239548, "compression_ratio": 1.2090032154340835, "no_speech_prob": 9.655437040179127e-12}, {"id": 4, "seek": 0, "start": 22.0, "end": 27.0, "text": "第一遍保证模式才是准确度的大头,准确度占比80%,", "tokens": [51465, 18049, 3330, 235, 24302, 5233, 223, 41908, 27584, 18888, 1541, 6336, 228, 38114, 106, 13127, 1546, 3582, 39862, 11, 6336, 228, 38114, 106, 13127, 5322, 254, 11706, 4702, 8923, 51715], "temperature": 0, "avg_logprob": -0.17544205622239548, "compression_ratio": 1.2090032154340835, "no_speech_prob": 9.655437040179127e-12}, {"id": 5, "seek": 2700, "start": 27.0, "end": 31.0, "text": "第二遍只是填充的,以弥补字幕不全。", "tokens": [50365, 19693, 3330, 235, 36859, 13331, 104, 2347, 227, 1546, 11, 3588, 5702, 98, 9890, 98, 22381, 40211, 1960, 11319, 1543, 50565], "temperature": 0, "avg_logprob": -0.0645239023872502, "compression_ratio": 1.4306569343065694, "no_speech_prob": 1.5578075288069648e-11}, {"id": 6, "seek": 2700, "start": 31.0, "end": 35.0, "text": "所以PAS1保证模式必选精细度最高的OutTalk加10。", "tokens": [50565, 7239, 47, 3160, 16, 24302, 5233, 223, 41908, 27584, 28531, 2215, 231, 34910, 10115, 228, 13127, 8661, 12979, 1546, 28353, 33210, 9990, 3279, 1543, 50765], "temperature": 0, "avg_logprob": -0.0645239023872502, "compression_ratio": 1.4306569343065694, "no_speech_prob": 1.5578075288069648e-11}, {"id": 7, "seek": 2700, "start": 35.0, "end": 43.0, "text": "而PAS2平衡模式准确度更低,仅仅为了填充,所以选精细度不高的Colero,保证模式更强。", "tokens": [50765, 11070, 47, 3160, 17, 16716, 9890, 94, 41908, 27584, 6336, 228, 38114, 106, 13127, 19002, 41377, 11, 1550, 227, 1550, 227, 13992, 2289, 13331, 104, 2347, 227, 11, 7239, 2215, 231, 34910, 10115, 228, 13127, 1960, 12979, 1546, 34, 401, 2032, 11, 24302, 5233, 223, 41908, 27584, 19002, 5702, 118, 1543, 51165], "temperature": 0, "avg_logprob": -0.0645239023872502, "compression_ratio": 1.4306569343065694, "no_speech_prob": 1.5578075288069648e-11}, {"id": 8, "seek": 2700, "start": 43.0, "end": 49.0, "text": "而且比平衡模式准确度强太多,保证模式是FP16,平衡模式是影的8。", "tokens": [51165, 22942, 11706, 16716, 9890, 94, 41908, 27584, 6336, 228, 38114, 106, 13127, 5702, 118, 9455, 6392, 11, 24302, 5233, 223, 41908, 27584, 1541, 45882, 6866, 11, 16716, 9890, 94, 41908, 27584, 1541, 16820, 1546, 23, 1543, 51465], "temperature": 0, "avg_logprob": -0.0645239023872502, "compression_ratio": 1.4306569343065694, "no_speech_prob": 1.5578075288069648e-11}, {"id": 9, "seek": 2700, "start": 49.0, "end": 56.0, "text": "所以必选以PAS1为主,PAS2填充,这样准确度更高后期人工校正更轻松。", "tokens": [51465, 7239, 28531, 2215, 231, 3588, 47, 3160, 16, 13992, 13557, 11, 47, 3160, 17, 13331, 104, 2347, 227, 11, 21209, 6336, 228, 38114, 106, 13127, 19002, 12979, 13547, 16786, 4035, 23323, 49684, 15789, 19002, 17819, 119, 36897, 1543, 51815], "temperature": 0, "avg_logprob": -0.0645239023872502, "compression_ratio": 1.4306569343065694, "no_speech_prob": 1.5578075288069648e-11}, {"id": 10, "seek": 5600, "start": 56.0, "end": 60.0, "text": "2. 模型必选Large VR而不是Large V3。", "tokens": [50365, 17, 13, 220, 41908, 39823, 28531, 2215, 231, 43, 289, 432, 13722, 11070, 7296, 43, 289, 432, 691, 18, 1543, 50565], "temperature": 0, "avg_logprob": -0.11092085790152502, "compression_ratio": 1.0604395604395604, "no_speech_prob": 1.4837608572348948e-11}, {"id": 11, "seek": 5600, "start": 60.0, "end": 66.0, "text": "VR等于日语专精型,V3等于全球平均型,转入日语首选V2。", "tokens": [50565, 44358, 10187, 37732, 6890, 5233, 255, 940, 241, 34910, 39823, 11, 53, 18, 10187, 37732, 11319, 28533, 16716, 14872, 229, 39823, 11, 17819, 105, 14028, 6890, 5233, 255, 25444, 2215, 231, 53, 17, 1543, 50865], "temperature": 0, "avg_logprob": -0.11092085790152502, "compression_ratio": 1.0604395604395604, "no_speech_prob": 1.4837608572348948e-11}, {"id": 12, "seek": 5600, "start": 66.0, "end": 72.0, "text": "V3做了两个大升级,梅尔平谱从80-128,日语更容易收噪音影响。", "tokens": [50865, 53, 18, 10907, 2289, 36257, 7549, 3582, 41670, 16853, 100, 11, 19017, 227, 1530, 242, 16716, 8897, 109, 35630, 4702, 12, 4762, 23, 11, 6890, 5233, 255, 19002, 49212, 18681, 21250, 103, 18034, 16820, 3858, 235, 1543, 51165], "temperature": 0, "avg_logprob": -0.11092085790152502, "compression_ratio": 1.0604395604395604, "no_speech_prob": 1.4837608572348948e-11}, {"id": 13, "seek": 7200, "start": 72.0, "end": 76.0, "text": "词调Token扩容加了很多小语种,日语全重下降。", "tokens": [50365, 5233, 235, 8897, 225, 51, 8406, 3416, 102, 25750, 9990, 2289, 20778, 7322, 5233, 255, 39810, 11, 6890, 5233, 255, 11319, 12624, 4438, 47421, 1543, 50565], "temperature": 0, "avg_logprob": -0.12332371798428622, "compression_ratio": 1.2041884816753927, "no_speech_prob": 1.7059229029392498e-11}, {"id": 14, "seek": 7200, "start": 76.0, "end": 79.0, "text": "这两个改动对日语全是浮悠化。", "tokens": [50565, 5562, 36257, 7549, 34490, 34961, 8713, 6890, 5233, 255, 11319, 1541, 11038, 106, 14696, 254, 23756, 1543, 50715], "temperature": 0, "avg_logprob": -0.12332371798428622, "compression_ratio": 1.2041884816753927, "no_speech_prob": 1.7059229029392498e-11}, {"id": 15, "seek": 7200, "start": 79.0, "end": 82.0, "text": "V3更改编,日语最受不了幻觉。", "tokens": [50715, 53, 18, 19002, 34490, 38109, 244, 11, 6890, 5233, 255, 8661, 23151, 47225, 3509, 119, 24447, 1543, 50865], "temperature": 0, "avg_logprob": -0.12332371798428622, "compression_ratio": 1.2041884816753927, "no_speech_prob": 1.7059229029392498e-11}, {"id": 16, "seek": 7200, "start": 82.0, "end": 89.0, "text": "V3为了流畅度更容易脑补内容,日语一脑补就是语法错,禁语错,注词。", "tokens": [50865, 53, 18, 13992, 2289, 27854, 6904, 227, 13127, 19002, 49212, 27067, 239, 9890, 98, 34742, 25750, 11, 6890, 5233, 255, 2257, 27067, 239, 9890, 98, 5620, 5233, 255, 11148, 29900, 11, 18614, 223, 5233, 255, 29900, 11, 26432, 5233, 235, 1543, 51215], "temperature": 0, "avg_logprob": -0.12332371798428622, "compression_ratio": 1.2041884816753927, "no_speech_prob": 1.7059229029392498e-11}, {"id": 17, "seek": 8900, "start": 89.0, "end": 93.0, "text": "错VR更保守,不乱编日语反而更稳。", "tokens": [50365, 29900, 44358, 19002, 24302, 48480, 11, 1960, 2930, 109, 38109, 244, 6890, 5233, 255, 22138, 11070, 19002, 10415, 111, 1543, 50565], "temperature": 0, "avg_logprob": -0.09482854207356771, "compression_ratio": 0.9575471698113207, "no_speech_prob": 1.549990517879518e-11}, {"id": 18, "seek": 8900, "start": 93.0, "end": 99.0, "text": "3. Whisper J A V尽管能把Large VR的幻觉降低,但是实测幻觉依然存在。", "tokens": [50565, 18, 13, 41132, 610, 508, 316, 691, 1530, 121, 23131, 8225, 16075, 43, 289, 432, 13722, 1546, 3509, 119, 24447, 47421, 41377, 11, 11189, 24726, 11038, 233, 3509, 119, 24447, 3254, 251, 5823, 39781, 3581, 1543, 50865], "temperature": 0, "avg_logprob": -0.09482854207356771, "compression_ratio": 0.9575471698113207, "no_speech_prob": 1.549990517879518e-11}, {"id": 19, "seek": 8900, "start": 99.0, "end": 104.0, "text": "可以使用AI翻译成中文后,用Subtitle Edit软件打开该字幕。", "tokens": [50865, 6723, 22982, 9254, 48698, 42716, 5233, 239, 11336, 5975, 17174, 13547, 11, 9254, 39582, 27689, 306, 33241, 17819, 107, 20485, 12467, 18937, 44646, 22381, 40211, 1543, 51115], "temperature": 0, "avg_logprob": -0.09482854207356771, "compression_ratio": 0.9575471698113207, "no_speech_prob": 1.549990517879518e-11}, {"id": 20, "seek": 10400, "start": 104.0, "end": 110.0, "text": "进入多重替换,用正则表达式或者正常字符替换添加幻觉词消除。", "tokens": [50365, 36700, 14028, 6392, 12624, 9531, 123, 26075, 95, 11, 9254, 15789, 2437, 247, 17571, 9830, 122, 27584, 31148, 15789, 11279, 22381, 5437, 99, 9531, 123, 26075, 95, 14176, 119, 9990, 3509, 119, 24447, 5233, 235, 28837, 32999, 1543, 50665], "temperature": 0, "avg_logprob": -0.06070480896876408, "compression_ratio": 1.2444444444444445, "no_speech_prob": 1.619619369175318e-11}, {"id": 21, "seek": 10400, "start": 110.0, "end": 116.0, "text": "Large VR转写加AI翻译后的幻觉并非随机乱码,而是有明确的高频固定模式。", "tokens": [50665, 43, 289, 432, 13722, 17819, 105, 5676, 247, 9990, 48698, 42716, 5233, 239, 13547, 1546, 3509, 119, 24447, 3509, 114, 12107, 10673, 237, 37960, 2930, 109, 23230, 223, 11, 11070, 1541, 2412, 11100, 38114, 106, 1546, 12979, 39752, 3919, 118, 12088, 41908, 27584, 1543, 50965], "temperature": 0, "avg_logprob": -0.06070480896876408, "compression_ratio": 1.2444444444444445, "no_speech_prob": 1.619619369175318e-11}, {"id": 22, "seek": 10400, "start": 116.0, "end": 121.0, "text": "多余的冗于双引号,莫名出现的人物前缀,固定同音措辞等。", "tokens": [50965, 6392, 1593, 247, 1546, 5676, 245, 37732, 2129, 234, 41301, 26987, 11, 31127, 104, 15940, 7781, 20204, 29979, 23516, 8945, 38109, 222, 11, 3919, 118, 12088, 13089, 18034, 6900, 103, 9830, 252, 10187, 1543, 51215], "temperature": 0, "avg_logprob": -0.06070480896876408, "compression_ratio": 1.2444444444444445, "no_speech_prob": 1.619619369175318e-11}, {"id": 23, "seek": 10400, "start": 121.0, "end": 127.0, "text": "色Title Edit的多重替换,支持一次性导入整份规则库,也就是我提到的幻觉词表。", "tokens": [51215, 17673, 51, 270, 306, 33241, 1546, 6392, 12624, 9531, 123, 26075, 95, 11, 35488, 27505, 21686, 4510, 120, 14028, 27662, 36266, 6758, 226, 2437, 247, 6346, 241, 11, 6404, 5620, 1654, 20949, 4511, 1546, 3509, 119, 24447, 5233, 235, 17571, 1543, 51515], "temperature": 0, "avg_logprob": -0.06070480896876408, "compression_ratio": 1.2444444444444445, "no_speech_prob": 1.619619369175318e-11}, {"id": 24, "seek": 10400, "start": 127.0, "end": 133.0, "text": "双引号修正规则,全局批量执行,还可以反复运行二到三次彻底清理。", "tokens": [51515, 2129, 234, 41301, 26987, 7792, 106, 15789, 6758, 226, 2437, 247, 11, 11319, 34703, 3416, 117, 26748, 3416, 100, 8082, 11, 14852, 6723, 22138, 1787, 235, 3316, 238, 8082, 11217, 4511, 10960, 9487, 7391, 119, 22816, 21784, 13876, 1543, 51815], "temperature": 0, "avg_logprob": -0.06070480896876408, "compression_ratio": 1.2444444444444445, "no_speech_prob": 1.619619369175318e-11}, {"id": 25, "seek": 13300, "start": 133.0, "end": 138.0, "text": "全程不需要逐行手动修改,效率远高于纯人工校对。", "tokens": [50365, 11319, 29649, 1960, 35748, 2215, 238, 8082, 11389, 34961, 7792, 106, 34490, 11, 43076, 44866, 3316, 250, 12979, 37732, 16853, 107, 4035, 23323, 49684, 8713, 1543, 50615], "temperature": 0, "avg_logprob": -0.07507155407434223, "compression_ratio": 1.1615853658536586, "no_speech_prob": 1.805357426054588e-11}, {"id": 26, "seek": 13300, "start": 138.0, "end": 142.0, "text": "Large VR的幻觉本质是训练数据高度依赖YouTube预料。", "tokens": [50615, 43, 289, 432, 13722, 1546, 3509, 119, 24447, 8802, 18464, 101, 1541, 7422, 255, 10115, 225, 33188, 26075, 106, 12979, 13127, 3254, 251, 5266, 244, 3223, 41173, 12501, 226, 33404, 1543, 50815], "temperature": 0, "avg_logprob": -0.07507155407434223, "compression_ratio": 1.1615853658536586, "no_speech_prob": 1.805357426054588e-11}, {"id": 27, "seek": 13300, "start": 142.0, "end": 149.0, "text": "模型在静音、噪音、低质性读音频段会自动脑补出训练里最高频的YouTube互动画书,", "tokens": [50815, 41908, 39823, 3581, 5363, 247, 18034, 1231, 21250, 103, 18034, 1231, 41377, 18464, 101, 21686, 5233, 119, 18034, 39752, 28427, 12949, 9722, 34961, 27067, 239, 9890, 98, 7781, 7422, 255, 10115, 225, 15759, 8661, 12979, 39752, 1546, 3223, 41173, 1369, 240, 34961, 27126, 2930, 99, 11, 51165], "temperature": 0, "avg_logprob": -0.07507155407434223, "compression_ratio": 1.1615853658536586, "no_speech_prob": 1.805357426054588e-11}, {"id": 28, "seek": 13300, "start": 149.0, "end": 152.0, "text": "属于模型先天特性,并非操作故障。", "tokens": [51165, 9636, 252, 37732, 41908, 39823, 10108, 6135, 17682, 21686, 11, 3509, 114, 12107, 13167, 235, 11914, 43045, 10673, 250, 1543, 51315], "temperature": 0, "avg_logprob": -0.07507155407434223, "compression_ratio": 1.1615853658536586, "no_speech_prob": 1.805357426054588e-11}, {"id": 29, "seek": 13300, "start": 152.0, "end": 159.0, "text": "4. 关于人工修正,经过色Pyto Edit清洗以上步骤后,修改起来还是比较轻松。", "tokens": [51315, 19, 13, 220, 28053, 37732, 4035, 23323, 7792, 106, 15789, 11, 30276, 16866, 17673, 47, 88, 1353, 33241, 21784, 43252, 29497, 31429, 49657, 97, 13547, 11, 7792, 106, 34490, 9147, 6912, 45726, 11706, 9830, 225, 17819, 119, 36897, 1543, 51665], "temperature": 0, "avg_logprob": -0.07507155407434223, "compression_ratio": 1.1615853658536586, "no_speech_prob": 1.805357426054588e-11}, {"id": 30, "seek": 15900, "start": 159.0, "end": 164.0, "text": "我愿称之为目前人工修正最轻松,幻觉最少的字幕软件。", "tokens": [50365, 1654, 4820, 123, 8204, 108, 9574, 13992, 39004, 4035, 23323, 7792, 106, 15789, 8661, 17819, 119, 36897, 11, 3509, 119, 24447, 8661, 15686, 1546, 22381, 40211, 17819, 107, 20485, 1543, 50615], "temperature": 0, "avg_logprob": -0.04769768127023357, "compression_ratio": 1.1137724550898203, "no_speech_prob": 2.1312866402278807e-11}, {"id": 31, "seek": 15900, "start": 164.0, "end": 172.0, "text": "一般300行5到10分钟左右,600行10到15分钟左右,快一点的话就5分钟,看你是不是想认真改。", "tokens": [50615, 2257, 49640, 12566, 8082, 20, 4511, 3279, 6627, 50064, 29457, 11, 15707, 8082, 3279, 4511, 5211, 6627, 50064, 29457, 11, 10251, 49042, 44575, 3111, 20, 6627, 50064, 11, 4200, 2166, 23034, 7093, 7422, 97, 6303, 34490, 1543, 51015], "temperature": 0, "avg_logprob": -0.04769768127023357, "compression_ratio": 1.1137724550898203, "no_speech_prob": 2.1312866402278807e-11}, {"id": 32, "seek": 17200, "start": 172.0, "end": 181.0, "text": "关于笔记本设备,推荐3070TI笔记本,拥有5888个流处理器,是性价比最高的AI设备选择。", "tokens": [50365, 28053, 37732, 5437, 242, 34756, 8802, 7422, 122, 1787, 229, 11, 33597, 31409, 238, 3446, 5867, 5422, 5437, 242, 34756, 8802, 11, 6852, 98, 2412, 20, 16919, 23, 7549, 27854, 1787, 226, 13876, 34386, 11, 1541, 21686, 1550, 115, 11706, 8661, 12979, 1546, 48698, 7422, 122, 1787, 229, 2215, 231, 6852, 102, 1543, 50815], "temperature": 0, "avg_logprob": -0.052726096465808, "compression_ratio": 1.1216216216216217, "no_speech_prob": 1.8018067940439586e-11}, {"id": 33, "seek": 17200, "start": 181.0, "end": 187.0, "text": "转录速度比4060亿4070万5060笔记本快,而且价格也不贵。", "tokens": [50815, 17819, 105, 7391, 243, 31217, 13127, 11706, 5254, 4550, 1369, 123, 5254, 5867, 23570, 2803, 4550, 5437, 242, 34756, 8802, 10251, 11, 22942, 1550, 115, 30921, 6404, 1960, 18464, 113, 1543, 51115], "temperature": 0, "avg_logprob": -0.052726096465808, "compression_ratio": 1.1216216216216217, "no_speech_prob": 1.8018067940439586e-11}, {"id": 34, "seek": 17200, "start": 187.0, "end": 192.0, "text": "只要把温度控制好,就能最大程度发挥效率,节省时间。", "tokens": [51115, 14003, 4275, 16075, 9592, 102, 13127, 48707, 25491, 2131, 11, 3111, 8225, 8661, 3582, 29649, 13127, 28926, 8501, 98, 43076, 44866, 11, 45161, 2862, 223, 44848, 1543, 51365], "temperature": 0, "avg_logprob": -0.052726096465808, "compression_ratio": 1.1216216216216217, "no_speech_prob": 1.8018067940439586e-11}, {"id": 35, "seek": 19200, "start": 192.0, "end": 196.0, "text": "5.字幕的质量还算可以,精确度非常高。", "tokens": [50365, 20, 13, 22381, 40211, 1546, 18464, 101, 26748, 14852, 19497, 6723, 11, 34910, 38114, 106, 13127, 14392, 12979, 1543, 50565], "temperature": 0, "avg_logprob": -0.05236226528674572, "compression_ratio": 1.25, "no_speech_prob": 1.7774658481184424e-11}, {"id": 36, "seek": 19200, "start": 196.0, "end": 203.0, "text": "当然尽管用了双通道模式避免漏掉字幕,但或多或少还是会漏掉一点字幕,不过瑕不掩瑜。", "tokens": [50565, 40486, 1530, 121, 23131, 9254, 2289, 2129, 234, 19550, 6025, 41908, 27584, 3330, 123, 2347, 235, 14065, 237, 29327, 22381, 40211, 11, 8395, 19780, 6392, 19780, 15686, 45726, 12949, 14065, 237, 29327, 49042, 22381, 40211, 11, 1960, 16866, 37955, 243, 1960, 6900, 102, 163, 12139, 1543, 50915], "temperature": 0, "avg_logprob": -0.05236226528674572, "compression_ratio": 1.25, "no_speech_prob": 1.7774658481184424e-11}, {"id": 37, "seek": 19200, "start": 203.0, "end": 208.0, "text": "如果强行上保真模式最高灵敏度,首先幻觉会直线上升。", "tokens": [50915, 13119, 5702, 118, 8082, 5708, 24302, 6303, 41908, 27584, 8661, 12979, 13290, 113, 7017, 237, 13127, 11, 36490, 3509, 119, 24447, 12949, 16186, 16853, 123, 5708, 41670, 1543, 51165], "temperature": 0, "avg_logprob": -0.05236226528674572, "compression_ratio": 1.25, "no_speech_prob": 1.7774658481184424e-11}, {"id": 38, "seek": 19200, "start": 208.0, "end": 215.0, "text": "其次是字幕人工修正更辛苦,两者直接影响后面的准确度,而且时间太漫长得不偿失。", "tokens": [51165, 9572, 9487, 1541, 22381, 40211, 4035, 23323, 7792, 106, 15789, 19002, 46472, 11, 36257, 12444, 43297, 16820, 3858, 235, 13547, 8833, 1546, 6336, 228, 38114, 106, 13127, 11, 22942, 44848, 9455, 14065, 104, 32271, 5916, 1960, 7437, 123, 28564, 1543, 51515], "temperature": 0, "avg_logprob": -0.05236226528674572, "compression_ratio": 1.25, "no_speech_prob": 1.7774658481184424e-11}, {"id": 39, "seek": 21500, "start": 215.0, "end": 220.0, "text": "为了这一点漏掉的字幕而放弃最终的准确度,没必要。", "tokens": [50365, 13992, 2289, 5562, 49042, 14065, 237, 29327, 1546, 22381, 40211, 11070, 12744, 5702, 225, 8661, 10115, 230, 1546, 6336, 228, 38114, 106, 13127, 11, 10062, 28531, 4275, 1543, 50615], "temperature": 0, "avg_logprob": -0.0828132943103188, "compression_ratio": 1.1904761904761905, "no_speech_prob": 1.688677322975174e-11}, {"id": 40, "seek": 21500, "start": 220.0, "end": 228.0, "text": "假如是一小时时评600行,Whisper这保真模式转录需要一小时20-30分钟左右,3070TI笔记本。", "tokens": [50615, 31706, 8238, 1541, 2257, 7322, 15729, 15729, 5233, 226, 15707, 8082, 11, 2471, 271, 610, 5562, 24302, 6303, 41908, 27584, 17819, 105, 7391, 243, 35748, 2257, 7322, 15729, 2009, 12, 3446, 6627, 50064, 29457, 11, 3446, 5867, 5422, 5437, 242, 34756, 8802, 1543, 51015], "temperature": 0, "avg_logprob": -0.0828132943103188, "compression_ratio": 1.1904761904761905, "no_speech_prob": 1.688677322975174e-11}, {"id": 41, "seek": 21500, "start": 228.0, "end": 232.0, "text": "AI翻译也就2-3分钟,多重替换也就10几秒。", "tokens": [51015, 48698, 42716, 5233, 239, 6404, 3111, 17, 12, 18, 6627, 50064, 11, 6392, 12624, 9531, 123, 26075, 95, 6404, 3111, 3279, 6336, 254, 20956, 1543, 51215], "temperature": 0, "avg_logprob": -0.0828132943103188, "compression_ratio": 1.1904761904761905, "no_speech_prob": 1.688677322975174e-11}, {"id": 42, "seek": 21500, "start": 232.0, "end": 234.0, "text": "人工校正一般10-15分钟。", "tokens": [51215, 4035, 23323, 49684, 15789, 2257, 49640, 3279, 12, 5211, 6627, 50064, 1543, 51315], "temperature": 0, "avg_logprob": -0.0828132943103188, "compression_ratio": 1.1904761904761905, "no_speech_prob": 1.688677322975174e-11}, {"id": 43, "seek": 21500, "start": 234.0, "end": 240.0, "text": "整体上看最耗费时间的还是转录,所以可以把转录时间留给凌晨。", "tokens": [51315, 27662, 29485, 5708, 4200, 8661, 4450, 245, 18464, 117, 44848, 1546, 45726, 17819, 105, 7391, 243, 11, 7239, 6723, 16075, 17819, 105, 7391, 243, 44848, 24456, 23197, 6336, 234, 5094, 101, 1543, 51615], "temperature": 0, "avg_logprob": -0.0828132943103188, "compression_ratio": 1.1904761904761905, "no_speech_prob": 1.688677322975174e-11}, {"id": 44, "seek": 24000, "start": 240.0, "end": 245.0, "text": "一般凌晨0点转录,到早上8点,一般就能转录好5个视频。", "tokens": [50365, 2257, 49640, 6336, 234, 5094, 101, 15, 12579, 17819, 105, 7391, 243, 11, 4511, 21176, 5708, 23, 12579, 11, 2257, 49640, 3111, 8225, 17819, 105, 7391, 243, 2131, 20, 7549, 40656, 39752, 1543, 50615], "temperature": 0, "avg_logprob": -0.06623730417025292, "compression_ratio": 1.0232558139534884, "no_speech_prob": 1.9228358488776465e-11}, {"id": 45, "seek": 24000, "start": 245.0, "end": 249.0, "text": "视频时长都是一小时左右,不影响白天玩电脑。", "tokens": [50615, 40656, 39752, 15729, 32271, 22796, 2257, 7322, 15729, 29457, 11, 1960, 16820, 3858, 235, 13558, 6135, 19912, 42182, 27067, 239, 1543, 50815], "temperature": 0, "avg_logprob": -0.06623730417025292, "compression_ratio": 1.0232558139534884, "no_speech_prob": 1.9228358488776465e-11}], "language": "zh"}