{
  "text": "大家好,今天给大家介绍一篇比较新的paper,也是基于Visper作语音识别的。这篇论文的题目叫Visper CD,专门解决长音频的语音识别问题。如果用一句话来总结的话,这篇paper它的核心亮点就在于不改变Visper模型的参数,也无需训练。只在Visper解码时给它加一个反向的揪片器,这样专门来压制长音频的语音识别。\n音频转写里面的幻觉\n重复循环和漏字等问题\n这就是这篇论文\n它核心要解决的问题\n这里面CD代表什么意思\n就是一个对比解码\n那我们下面来看一下\n这篇PayPal它主要解决什么问题\n它核心要解决的\n就是可能比较长的一些语音\n然后呢\n以Visper为代表的一些语音识别模型\n它面临的幻觉\n重复循环和内容遗漏\n等错误累积的问题\n因为上海文太长了嘛\n而且这种Visper\n它是自回归的一个模型结构\n所以上文的错误会累积到下文\n这三种错误基本上可以看成分成以下三类\n第一种就是静音区的幻觉\n例如在非语音的间隔期间\n但是模型呢\n它还是错误的生成了文本\n第二个就是跨段的重复的循环\n就是模型陷入自我强化的重复模式\n它会错误的不停的输出相同的字\n然后一直循环一直循环\n然后永远不停\n第三个就是跳过\n那就是遗漏了实际的语音内容\n导致转路不完整\n实际上就是我们语音识别里常见的三类错误\n插入删除替换这三类错误\n此外这个论文也指出\n当前因为是自回归的语音识别模型\n它会利用前一段转路的结果作为解码的上下文\n那么这样就会带来一个问题\n就是上面的错误会被累积\n引起后面解码带来更多的错误\n这就要形成了新的累积错误\n这也是自卫规模型它常见的问题\n那这篇paper它是如何解决这个问题的呢\n它就提出了叫Visper CD\n这种方式啊\n不需要训练这个Visper模型的\n在推理时对比干净的音频和声学扰动后的音频的一个输出\n一个logis\n然后从而呢\n抑制长形式语音转换中的一些插入删除和替换错误\n那我们就来看一下它具体是怎么做的\n我们先看右面这张图吧\n我把这张图看懂了\n那这篇论文最精华最核心的部分也就懂了\n首先白色这一部分就是输入的语音\n然后这一部分呢\n就是粉色的这部分就是Whisper\nWhisper我们知道\n它就整体过程中就不需要动\n那过程是怎么样的呢\n我们先看虚线这部分啊\n就是上一个30秒已经识别完\n识别之后的一个语音片段\n那现在呢\n我们要识别最新的\n这个新的这30秒的语音片段\n需要把它给到我们Whisper的encoder\n那你看这就是黑色这部分\n就是原始的这段\n此外呢\n我们会对原始的音频\n做一些相关的扰动\n把它当做复理\n原来这段就当做正理\n复理第一个种复理呢\n就是我们加一些高斯噪声\n就是在原始的这段干净音频上\n加一些高斯噪声\n把它变成第一种复理\n也就是这个蓝色这部分\n第二种复利呢\n就是我们把原始的音频啊\n向右稍微偏移一小段\n这个就变成shift\n就是偏移的这段扰动的音频\n第三种音频啊\n就是silence\n就是完全静音的这段音频\n然后呢\n我们把这四种音频\n同时送给Whisper\n它的encoder和decoder去解码\n解码之后呢\n我们就得到了对应的四个输出\n那黑色这部分就是原始的输出\n蓝色绿色和红色分别也对应了三个输出\n然后呢\n我们就把正力的得分输出和负力的得分输出\n进行一个整合\n以叫我们叫对比的一个解码\n解码之后就可以帮助我们得到\n更准确的语音识别的结果\n那具体它是怎么去做这个对比解码的呢\n我们就来看一下论文中的原始\n它是怎么说的\n其实方式也很简单\n正力得到的这个X就是干净的\n原始干净的语音\n那自回归之后\n得到我们正力的一个结果\n复利的呢\n这个X上面加一个波浪\n就是我们经过三种变换\n加高斯啊\n加偏移啊\n和静音的变换之后的这种音频\n也经过我们F\n也就是Whisper之后\n得到我们复利的输出\n然后呢\n我就通过一个α的一个超三\n把这两种得分输出\n把它整合在一起\n然后就变成我们最终的一个CD的一个输出\n那具体是怎么整合的呢\n实际上就是1加上α\n然后乘以我们正力的输出\n减去α乘以我们负力的输出\n实际上你如果把这个式子变换一下\n其实就是把这个括号拆开啊\n就变成正力的输出\n加上正力的输出\n减去负力的输出\n然后再乘一个alpha的参数\n对吧\n所以这它代表的是什么含义呢\n实际上就是我在保持正力输出不变的前提下\n我把负力的那些扰动\n希望都把它剪掉\n然后把这些干扰项去掉\n然后让它有一个对比的结果\n从而帮助我更好的识别\n最终的语音识别结果\n这就是这篇paper最重要的\n也最核心的一个机制\n叫对比解码\n那下面我们就来看一下\n它声学扰动具体是怎么做的\n基本上就是三种\n一个是高斯造声的注入\n就是我们随机生成\n符合正态分布的高斯造声\n加到原始信号上\n第二个呢就是静音\n我就把输入图谱\n全都变成零\n全零的波形\n就是静音\n第三个呢就是音频的时序偏移\n我就把原始的30秒\n向左偏移一些\n然后左侧的丢弃\n右侧的用零来填充\n这就好了\n这就是三种升学扰动的策略\n然后呢\n如果我们有了升学扰动的策略\n那这三种策略\n它是怎么结合在一起的呢\n结合其实也非常简单\n大家还记得我原始的诗子吧\n就是1加上α乘以正力\n减去α乘以负力\n这里只不过负力有很多啊\n就是有三种扰动策略嘛\n所以负力产生的输出结果就有三种\n那这三种我怎么整合到一起呢\n我就把它在e的指数上\n进行一个整合\n先求和再求个平均\n无非是这里加了一个超次参数\n来控制一下\n对吧\n所以这个超次参数呢\n一般也就设为1\n那这个1其实就等于没有吗\n你除它就等于没有了\n那其实在原始的例子上\n无非就是把三个负利的输出\n进行求和再取平均\n就非常简单吧\n然后呢,我们再看在推理的时候。推理的时候,我们知道,呃,其实是可以并行推理的嘛,对吧?就是尽管输入的是K条额外的解码路径,但是我们可以通过P处理来计算控制的开销,也就是一次处理把它\n都做完\n然后整体的\n它最大的优势\n就是无需训练\n然后我只需要在后面\n只需要在解码的时候\n稍加一些\n所谓的工程化的处理就够了\n那我们可以在这里\n做一个形象的比喻\n就是如果你上面\n听的还不是特别懂的话\n如果我们把原始的Whisper\n它正常转型的时候\n它就是像一个\n边听边写的速记员\n那这篇paper\n它这个Whisper CD\n它实际上做的就是\n在旁边放了三个捣乱版的监听器\n那一个呢\n让它听起来很吵\n也就是加噪音的\n加高斯噪声的\n一个让它几乎什么都听不到\n然后就是静音的\n还有一个呢\n是让它错位的\n就是这个偏移的\n然后如果某个词\n在这三种不利的条件下\n它仍然特别想写出来\n那这个词多半就不是真的想听到的\n而是它习惯性脑补的\n于是系统就把这些词的分数扣掉\n然后这个解释啊\n基本上或者这个比喻啊\n就是整篇论文方法的本质\n嗯\n最后我们来看一下它的效果啊\n整个论文呢\n是在这五个长音频的英文测试机上来做的\n然后我们看一下\n它这个效果还是很不错的\n就是加CD是这次改编之后的吧\n然后它在两个VSPR的版本上做\n我们看到原始的VSPR\n像V3那种\n它在这些测试机上\n这个凑率都非常高\n这都超过100%了\n有200多500多\n但是呢我们在Visper V3上加了解码上的改进之后\n下降非常多\n就都从200多变成45\n从44变成19\n从500多变成57\n对吧\n第二种方式呢\n就是我用更强的版本\n用Turbo这个版本\n虽然原始的版本\n直接用Visper\n识别这个长文效果比V3会好一些啊\n但是也不尽如人意吧\n但是呢\n我们加了这种CD之后\n你会看到它效果像样的也是非常明显\n所以这就是我觉得这篇paper非常\n值得分享的吧\n就是它用非常小的改动\n然后实现了对常文的一个效果非常大的提升\n我们最后来总结一下这篇paper\n就是长音频的识别\nWhisper的话会有很多严重的错误\n但这些错误它大概率不是模型不会识别\n而是解码的过程中\n被之前的识别结果和一些上下文绑架了\n导致它越识别越错\n越识别越错\n所以作者他不是去强化原始模型的能力\n也就是没有训练原始的模型\n而是直接在token选择的阶段做干预\n这个角度就是在工程测作\n然后也很聪明\n好这期视频就给大家介绍到这",
  "language": "zh",
  "duration": 605.947938,
  "segments": [
    {
      "start": 0.0,
      "end": 29.98,
      "text": "大家好,今天给大家介绍一篇比较新的paper,也是基于Visper作语音识别的。这篇论文的题目叫Visper CD,专门解决长音频的语音识别问题。如果用一句话来总结的话,这篇paper它的核心亮点就在于不改变Visper模型的参数,也无需训练。只在Visper解码时给它加一个反向的揪片器,这样专门来压制长音频的语音识别。",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.16168492879623023,
      "no_speech_prob": 1.0607980639731895e-11,
      "compression_ratio": 1.3367003367003367
    },
    {
      "start": 30.0,
      "end": 31.74,
      "text": "音频转写里面的幻觉",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 31.74,
      "end": 34.58,
      "text": "重复循环和漏字等问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 34.58,
      "end": 36.5,
      "text": "这就是这篇论文",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 36.5,
      "end": 37.92,
      "text": "它核心要解决的问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 37.92,
      "end": 39.36,
      "text": "这里面CD代表什么意思",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 39.36,
      "end": 40.42,
      "text": "就是一个对比解码",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 40.42,
      "end": 41.54,
      "text": "那我们下面来看一下",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 41.54,
      "end": 43.36,
      "text": "这篇PayPal它主要解决什么问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 43.36,
      "end": 45.04,
      "text": "它核心要解决的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 45.04,
      "end": 47.32,
      "text": "就是可能比较长的一些语音",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 47.32,
      "end": 48.06,
      "text": "然后呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 48.06,
      "end": 50.68,
      "text": "以Visper为代表的一些语音识别模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 50.68,
      "end": 52.28,
      "text": "它面临的幻觉",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 52.28,
      "end": 55.22,
      "text": "重复循环和内容遗漏",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 55.22,
      "end": 56.5,
      "text": "等错误累积的问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 56.5,
      "end": 58.08,
      "text": "因为上海文太长了嘛",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 58.08,
      "end": 59.34,
      "text": "而且这种Visper",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1076267690074687,
      "no_speech_prob": 3.1674517175783734e-11,
      "compression_ratio": 1.4483870967741936
    },
    {
      "start": 59.34,
      "end": 62.3,
      "text": "它是自回归的一个模型结构",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 62.3,
      "end": 64.54,
      "text": "所以上文的错误会累积到下文",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 64.54,
      "end": 68.78,
      "text": "这三种错误基本上可以看成分成以下三类",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 68.78,
      "end": 71.1,
      "text": "第一种就是静音区的幻觉",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 71.1,
      "end": 74.32,
      "text": "例如在非语音的间隔期间",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 74.32,
      "end": 75.4,
      "text": "但是模型呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 75.4,
      "end": 77.2,
      "text": "它还是错误的生成了文本",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 77.2,
      "end": 80.8,
      "text": "第二个就是跨段的重复的循环",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 80.8,
      "end": 83.26,
      "text": "就是模型陷入自我强化的重复模式",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 83.26,
      "end": 86.32,
      "text": "它会错误的不停的输出相同的字",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 86.32,
      "end": 87.76,
      "text": "然后一直循环一直循环",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 87.76,
      "end": 89.1,
      "text": "然后永远不停",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07112087444825606,
      "no_speech_prob": 1.3949563726356473e-11,
      "compression_ratio": 1.4529616724738676
    },
    {
      "start": 89.1,
      "end": 90.74,
      "text": "第三个就是跳过",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 90.74,
      "end": 93.96,
      "text": "那就是遗漏了实际的语音内容",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 93.96,
      "end": 95.3,
      "text": "导致转路不完整",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 95.3,
      "end": 98.36,
      "text": "实际上就是我们语音识别里常见的三类错误",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 98.36,
      "end": 101.0,
      "text": "插入删除替换这三类错误",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 101.0,
      "end": 102.92,
      "text": "此外这个论文也指出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 102.92,
      "end": 106.38,
      "text": "当前因为是自回归的语音识别模型",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 106.38,
      "end": 110.38,
      "text": "它会利用前一段转路的结果作为解码的上下文",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 110.38,
      "end": 112.78,
      "text": "那么这样就会带来一个问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 112.78,
      "end": 114.66,
      "text": "就是上面的错误会被累积",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 114.66,
      "end": 117.14,
      "text": "引起后面解码带来更多的错误",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.07623748029215952,
      "no_speech_prob": 1.4178356405603054e-11,
      "compression_ratio": 1.4270833333333333
    },
    {
      "start": 117.14,
      "end": 119.18,
      "text": "这就要形成了新的累积错误",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 119.18,
      "end": 121.68,
      "text": "这也是自卫规模型它常见的问题",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 121.68,
      "end": 124.76,
      "text": "那这篇paper它是如何解决这个问题的呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 124.76,
      "end": 126.66,
      "text": "它就提出了叫Visper CD",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 126.66,
      "end": 129.0,
      "text": "这种方式啊",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 129.0,
      "end": 130.96,
      "text": "不需要训练这个Visper模型的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 130.96,
      "end": 136.46,
      "text": "在推理时对比干净的音频和声学扰动后的音频的一个输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 136.46,
      "end": 137.42,
      "text": "一个logis",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 137.42,
      "end": 138.56,
      "text": "然后从而呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 138.56,
      "end": 143.26,
      "text": "抑制长形式语音转换中的一些插入删除和替换错误",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 143.26,
      "end": 146.0,
      "text": "那我们就来看一下它具体是怎么做的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1231268721017224,
      "no_speech_prob": 1.4113910561108778e-11,
      "compression_ratio": 1.2874617737003058
    },
    {
      "start": 146.0,
      "end": 148.38,
      "text": "我们先看右面这张图吧",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 148.38,
      "end": 149.82,
      "text": "我把这张图看懂了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 149.82,
      "end": 153.08,
      "text": "那这篇论文最精华最核心的部分也就懂了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 153.08,
      "end": 156.44,
      "text": "首先白色这一部分就是输入的语音",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 156.44,
      "end": 157.86,
      "text": "然后这一部分呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 157.86,
      "end": 159.98,
      "text": "就是粉色的这部分就是Whisper",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 159.98,
      "end": 161.56,
      "text": "Whisper我们知道",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 161.56,
      "end": 164.14,
      "text": "它就整体过程中就不需要动",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 164.14,
      "end": 165.86,
      "text": "那过程是怎么样的呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 165.86,
      "end": 167.36,
      "text": "我们先看虚线这部分啊",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 167.36,
      "end": 170.34,
      "text": "就是上一个30秒已经识别完",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 170.34,
      "end": 172.5,
      "text": "识别之后的一个语音片段",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 172.5,
      "end": 173.52,
      "text": "那现在呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09874483539212135,
      "no_speech_prob": 1.5293670843630203e-11,
      "compression_ratio": 1.4140350877192982
    },
    {
      "start": 173.52,
      "end": 174.86,
      "text": "我们要识别最新的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 174.86,
      "end": 176.9,
      "text": "这个新的这30秒的语音片段",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 176.9,
      "end": 180.82,
      "text": "需要把它给到我们Whisper的encoder",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 180.82,
      "end": 182.74,
      "text": "那你看这就是黑色这部分",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 182.74,
      "end": 184.28,
      "text": "就是原始的这段",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 184.28,
      "end": 185.56,
      "text": "此外呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 185.56,
      "end": 187.62,
      "text": "我们会对原始的音频",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 187.62,
      "end": 189.24,
      "text": "做一些相关的扰动",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 189.24,
      "end": 190.58,
      "text": "把它当做复理",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 190.58,
      "end": 192.42,
      "text": "原来这段就当做正理",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 192.42,
      "end": 194.26,
      "text": "复理第一个种复理呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 194.26,
      "end": 196.18,
      "text": "就是我们加一些高斯噪声",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 196.18,
      "end": 198.28,
      "text": "就是在原始的这段干净音频上",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 198.28,
      "end": 199.44,
      "text": "加一些高斯噪声",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 199.44,
      "end": 201.02,
      "text": "把它变成第一种复理",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 201.02,
      "end": 202.6,
      "text": "也就是这个蓝色这部分",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10107437675521219,
      "no_speech_prob": 1.473224840731202e-11,
      "compression_ratio": 1.5460992907801419
    },
    {
      "start": 202.6,
      "end": 204.0,
      "text": "第二种复利呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 204.0,
      "end": 206.14,
      "text": "就是我们把原始的音频啊",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 206.14,
      "end": 208.18,
      "text": "向右稍微偏移一小段",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 208.18,
      "end": 210.2,
      "text": "这个就变成shift",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 210.2,
      "end": 213.1,
      "text": "就是偏移的这段扰动的音频",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 213.1,
      "end": 214.5,
      "text": "第三种音频啊",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 214.5,
      "end": 215.76,
      "text": "就是silence",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 215.76,
      "end": 217.52,
      "text": "就是完全静音的这段音频",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 217.52,
      "end": 218.46,
      "text": "然后呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 218.46,
      "end": 219.96,
      "text": "我们把这四种音频",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 219.96,
      "end": 221.32,
      "text": "同时送给Whisper",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 221.32,
      "end": 223.76,
      "text": "它的encoder和decoder去解码",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 223.76,
      "end": 225.62,
      "text": "解码之后呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 225.62,
      "end": 228.86,
      "text": "我们就得到了对应的四个输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 228.86,
      "end": 231.26,
      "text": "那黑色这部分就是原始的输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09684565176371655,
      "no_speech_prob": 1.386873602071681e-11,
      "compression_ratio": 1.4204545454545454
    },
    {
      "start": 231.26,
      "end": 235.6,
      "text": "蓝色绿色和红色分别也对应了三个输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 235.6,
      "end": 236.42,
      "text": "然后呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 236.42,
      "end": 239.96,
      "text": "我们就把正力的得分输出和负力的得分输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 239.96,
      "end": 241.14,
      "text": "进行一个整合",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 241.14,
      "end": 243.34,
      "text": "以叫我们叫对比的一个解码",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 243.34,
      "end": 247.42,
      "text": "解码之后就可以帮助我们得到",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 247.42,
      "end": 249.98,
      "text": "更准确的语音识别的结果",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 249.98,
      "end": 253.12,
      "text": "那具体它是怎么去做这个对比解码的呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 253.12,
      "end": 256.06,
      "text": "我们就来看一下论文中的原始",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 256.06,
      "end": 257.04,
      "text": "它是怎么说的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 257.04,
      "end": 258.82,
      "text": "其实方式也很简单",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10897955989206073,
      "no_speech_prob": 1.0283108566044863e-11,
      "compression_ratio": 1.4312977099236641
    },
    {
      "start": 258.82,
      "end": 262.48,
      "text": "正力得到的这个X就是干净的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 262.48,
      "end": 264.02,
      "text": "原始干净的语音",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 264.02,
      "end": 266.38,
      "text": "那自回归之后",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 266.38,
      "end": 268.36,
      "text": "得到我们正力的一个结果",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 268.36,
      "end": 269.32,
      "text": "复利的呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 269.32,
      "end": 271.52,
      "text": "这个X上面加一个波浪",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 271.52,
      "end": 273.78,
      "text": "就是我们经过三种变换",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 273.78,
      "end": 274.92,
      "text": "加高斯啊",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 274.92,
      "end": 276.34,
      "text": "加偏移啊",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 276.34,
      "end": 279.24,
      "text": "和静音的变换之后的这种音频",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 279.24,
      "end": 281.06,
      "text": "也经过我们F",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 281.06,
      "end": 283.48,
      "text": "也就是Whisper之后",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 283.48,
      "end": 285.04,
      "text": "得到我们复利的输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 285.04,
      "end": 286.08,
      "text": "然后呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.14814044407435825,
      "no_speech_prob": 1.2270024406235702e-11,
      "compression_ratio": 1.3920704845814977
    },
    {
      "start": 286.08,
      "end": 288.86,
      "text": "我就通过一个α的一个超三",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 288.86,
      "end": 291.26,
      "text": "把这两种得分输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 291.26,
      "end": 292.72,
      "text": "把它整合在一起",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 292.72,
      "end": 296.62,
      "text": "然后就变成我们最终的一个CD的一个输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 296.62,
      "end": 298.62,
      "text": "那具体是怎么整合的呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 298.62,
      "end": 301.08,
      "text": "实际上就是1加上α",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 301.08,
      "end": 303.12,
      "text": "然后乘以我们正力的输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 303.12,
      "end": 305.4,
      "text": "减去α乘以我们负力的输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 305.4,
      "end": 308.02,
      "text": "实际上你如果把这个式子变换一下",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 308.02,
      "end": 310.76,
      "text": "其实就是把这个括号拆开啊",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 310.76,
      "end": 312.14,
      "text": "就变成正力的输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 312.14,
      "end": 314.26,
      "text": "加上正力的输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 314.26,
      "end": 315.48,
      "text": "减去负力的输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.09540698311545633,
      "no_speech_prob": 1.2560815235151956e-11,
      "compression_ratio": 1.6341463414634145
    },
    {
      "start": 315.48,
      "end": 317.16,
      "text": "然后再乘一个alpha的参数",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 317.16,
      "end": 317.6,
      "text": "对吧",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 317.6,
      "end": 321.0,
      "text": "所以这它代表的是什么含义呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 321.0,
      "end": 324.8,
      "text": "实际上就是我在保持正力输出不变的前提下",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 324.8,
      "end": 327.1,
      "text": "我把负力的那些扰动",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 327.1,
      "end": 328.44,
      "text": "希望都把它剪掉",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 328.44,
      "end": 330.4,
      "text": "然后把这些干扰项去掉",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 330.4,
      "end": 332.02,
      "text": "然后让它有一个对比的结果",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 332.02,
      "end": 333.98,
      "text": "从而帮助我更好的识别",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 333.98,
      "end": 336.06,
      "text": "最终的语音识别结果",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 336.06,
      "end": 339.48,
      "text": "这就是这篇paper最重要的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 339.48,
      "end": 341.2,
      "text": "也最核心的一个机制",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 341.2,
      "end": 342.02,
      "text": "叫对比解码",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 342.02,
      "end": 343.8,
      "text": "那下面我们就来看一下",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.1052978354764272,
      "no_speech_prob": 1.156865402085483e-11,
      "compression_ratio": 1.3236245954692556
    },
    {
      "start": 343.8,
      "end": 346.22,
      "text": "它声学扰动具体是怎么做的",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 346.22,
      "end": 347.34,
      "text": "基本上就是三种",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 347.34,
      "end": 349.0,
      "text": "一个是高斯造声的注入",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 349.0,
      "end": 350.44,
      "text": "就是我们随机生成",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 350.44,
      "end": 352.44,
      "text": "符合正态分布的高斯造声",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 352.44,
      "end": 353.74,
      "text": "加到原始信号上",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 353.74,
      "end": 355.48,
      "text": "第二个呢就是静音",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 355.48,
      "end": 356.82,
      "text": "我就把输入图谱",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 356.82,
      "end": 358.18,
      "text": "全都变成零",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 358.18,
      "end": 359.18,
      "text": "全零的波形",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 359.18,
      "end": 360.16,
      "text": "就是静音",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 360.16,
      "end": 362.56,
      "text": "第三个呢就是音频的时序偏移",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 362.56,
      "end": 364.68,
      "text": "我就把原始的30秒",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 364.68,
      "end": 367.18,
      "text": "向左偏移一些",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 367.18,
      "end": 369.08,
      "text": "然后左侧的丢弃",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 369.08,
      "end": 370.76,
      "text": "右侧的用零来填充",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 370.76,
      "end": 371.62,
      "text": "这就好了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10407006055459209,
      "no_speech_prob": 1.5205618014713096e-11,
      "compression_ratio": 1.424908424908425
    },
    {
      "start": 371.62,
      "end": 374.26,
      "text": "这就是三种升学扰动的策略",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 374.26,
      "end": 375.54,
      "text": "然后呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 375.54,
      "end": 377.76,
      "text": "如果我们有了升学扰动的策略",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 377.76,
      "end": 378.98,
      "text": "那这三种策略",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 378.98,
      "end": 381.26,
      "text": "它是怎么结合在一起的呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 381.26,
      "end": 382.68,
      "text": "结合其实也非常简单",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 382.68,
      "end": 384.72,
      "text": "大家还记得我原始的诗子吧",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 384.72,
      "end": 387.0,
      "text": "就是1加上α乘以正力",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 387.0,
      "end": 388.48,
      "text": "减去α乘以负力",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 388.48,
      "end": 390.36,
      "text": "这里只不过负力有很多啊",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 390.36,
      "end": 392.74,
      "text": "就是有三种扰动策略嘛",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 392.74,
      "end": 396.36,
      "text": "所以负力产生的输出结果就有三种",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 396.36,
      "end": 398.52,
      "text": "那这三种我怎么整合到一起呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 398.52,
      "end": 400.94,
      "text": "我就把它在e的指数上",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.10283242361886161,
      "no_speech_prob": 1.4452547665577686e-11,
      "compression_ratio": 1.4893617021276595
    },
    {
      "start": 400.94,
      "end": 401.96,
      "text": "进行一个整合",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 401.96,
      "end": 403.76,
      "text": "先求和再求个平均",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 403.76,
      "end": 406.28,
      "text": "无非是这里加了一个超次参数",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 406.28,
      "end": 407.28,
      "text": "来控制一下",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 407.28,
      "end": 407.7,
      "text": "对吧",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 407.7,
      "end": 409.2,
      "text": "所以这个超次参数呢",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 409.2,
      "end": 410.2,
      "text": "一般也就设为1",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 410.2,
      "end": 413.0,
      "text": "那这个1其实就等于没有吗",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 413.0,
      "end": 414.1,
      "text": "你除它就等于没有了",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 414.1,
      "end": 416.08,
      "text": "那其实在原始的例子上",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 416.08,
      "end": 419.6,
      "text": "无非就是把三个负利的输出",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 419.6,
      "end": 421.56,
      "text": "进行求和再取平均",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 421.56,
      "end": 422.7,
      "text": "就非常简单吧",
      "chunk": 0,
      "language": "zh",
      "avg_logprob": -0.13862077579941862,
      "no_speech_prob": 1.708289586177525e-11,
      "compression_ratio": 1.3842794759825328
    },
    {
      "start": 423.927,
      "end": 438.207,
      "text": "然后呢,我们再看在推理的时候。推理的时候,我们知道,呃,其实是可以并行推理的嘛,对吧?就是尽管输入的是K条额外的解码路径,但是我们可以通过P处理来计算控制的开销,也就是一次处理把它",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.19481385119853575,
      "no_speech_prob": 9.777175596914489e-12,
      "compression_ratio": 1.404191616766467
    },
    {
      "start": 438.207,
      "end": 438.927,
      "text": "都做完",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 438.927,
      "end": 439.947,
      "text": "然后整体的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 439.947,
      "end": 440.887,
      "text": "它最大的优势",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 440.887,
      "end": 441.867,
      "text": "就是无需训练",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 441.867,
      "end": 444.047,
      "text": "然后我只需要在后面",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 444.047,
      "end": 446.467,
      "text": "只需要在解码的时候",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 446.467,
      "end": 447.487,
      "text": "稍加一些",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 447.487,
      "end": 450.427,
      "text": "所谓的工程化的处理就够了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 450.427,
      "end": 452.687,
      "text": "那我们可以在这里",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 452.687,
      "end": 453.867,
      "text": "做一个形象的比喻",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 453.867,
      "end": 454.927,
      "text": "就是如果你上面",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 454.927,
      "end": 456.827,
      "text": "听的还不是特别懂的话",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 456.827,
      "end": 459.307,
      "text": "如果我们把原始的Whisper",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 459.307,
      "end": 460.647,
      "text": "它正常转型的时候",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 460.647,
      "end": 461.567,
      "text": "它就是像一个",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 461.567,
      "end": 463.007,
      "text": "边听边写的速记员",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 463.007,
      "end": 464.867,
      "text": "那这篇paper",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 464.867,
      "end": 465.987,
      "text": "它这个Whisper CD",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 465.987,
      "end": 467.667,
      "text": "它实际上做的就是",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.13260732867760572,
      "no_speech_prob": 2.2270074673258478e-11,
      "compression_ratio": 1.3474025974025974
    },
    {
      "start": 467.667,
      "end": 470.727,
      "text": "在旁边放了三个捣乱版的监听器",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 470.727,
      "end": 471.767,
      "text": "那一个呢",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 471.767,
      "end": 472.787,
      "text": "让它听起来很吵",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 472.787,
      "end": 474.307,
      "text": "也就是加噪音的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 474.307,
      "end": 475.607,
      "text": "加高斯噪声的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 475.607,
      "end": 477.647,
      "text": "一个让它几乎什么都听不到",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 477.647,
      "end": 480.147,
      "text": "然后就是静音的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 480.147,
      "end": 481.027,
      "text": "还有一个呢",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 481.027,
      "end": 482.847,
      "text": "是让它错位的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 482.847,
      "end": 484.107,
      "text": "就是这个偏移的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 484.107,
      "end": 485.467,
      "text": "然后如果某个词",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 485.467,
      "end": 487.247,
      "text": "在这三种不利的条件下",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 487.247,
      "end": 489.107,
      "text": "它仍然特别想写出来",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 489.107,
      "end": 492.927,
      "text": "那这个词多半就不是真的想听到的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 492.927,
      "end": 494.467,
      "text": "而是它习惯性脑补的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 494.467,
      "end": 497.007,
      "text": "于是系统就把这些词的分数扣掉",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08920181062486436,
      "no_speech_prob": 1.61934771147898e-11,
      "compression_ratio": 1.4183673469387754
    },
    {
      "start": 497.007,
      "end": 498.707,
      "text": "然后这个解释啊",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 498.707,
      "end": 499.987,
      "text": "基本上或者这个比喻啊",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 499.987,
      "end": 502.127,
      "text": "就是整篇论文方法的本质",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 502.127,
      "end": 502.867,
      "text": "嗯",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 502.867,
      "end": 505.287,
      "text": "最后我们来看一下它的效果啊",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 505.287,
      "end": 505.947,
      "text": "整个论文呢",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 505.947,
      "end": 509.707,
      "text": "是在这五个长音频的英文测试机上来做的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 509.707,
      "end": 510.647,
      "text": "然后我们看一下",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 510.647,
      "end": 513.407,
      "text": "它这个效果还是很不错的",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 513.407,
      "end": 516.347,
      "text": "就是加CD是这次改编之后的吧",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 516.347,
      "end": 518.667,
      "text": "然后它在两个VSPR的版本上做",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 518.667,
      "end": 520.787,
      "text": "我们看到原始的VSPR",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 520.787,
      "end": 522.507,
      "text": "像V3那种",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 522.507,
      "end": 523.887,
      "text": "它在这些测试机上",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 523.887,
      "end": 525.707,
      "text": "这个凑率都非常高",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.11724232164628666,
      "no_speech_prob": 1.3957729937119634e-11,
      "compression_ratio": 1.431578947368421
    },
    {
      "start": 525.707,
      "end": 527.427,
      "text": "这都超过100%了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 527.427,
      "end": 528.907,
      "text": "有200多500多",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 528.907,
      "end": 533.847,
      "text": "但是呢我们在Visper V3上加了解码上的改进之后",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 533.847,
      "end": 534.947,
      "text": "下降非常多",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 534.947,
      "end": 537.027,
      "text": "就都从200多变成45",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 537.027,
      "end": 538.647,
      "text": "从44变成19",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 538.647,
      "end": 539.827,
      "text": "从500多变成57",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 539.827,
      "end": 540.267,
      "text": "对吧",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 540.267,
      "end": 541.607,
      "text": "第二种方式呢",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 541.607,
      "end": 542.927,
      "text": "就是我用更强的版本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 542.927,
      "end": 544.007,
      "text": "用Turbo这个版本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 544.007,
      "end": 545.827,
      "text": "虽然原始的版本",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 545.827,
      "end": 547.647,
      "text": "直接用Visper",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 547.647,
      "end": 551.587,
      "text": "识别这个长文效果比V3会好一些啊",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 551.587,
      "end": 554.067,
      "text": "但是也不尽如人意吧",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 554.067,
      "end": 554.967,
      "text": "但是呢",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.18323025857248615,
      "no_speech_prob": 1.3062956098763845e-11,
      "compression_ratio": 1.2375886524822695
    },
    {
      "start": 554.967,
      "end": 556.467,
      "text": "我们加了这种CD之后",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 556.467,
      "end": 559.027,
      "text": "你会看到它效果像样的也是非常明显",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 559.027,
      "end": 561.747,
      "text": "所以这就是我觉得这篇paper非常",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 561.747,
      "end": 564.507,
      "text": "值得分享的吧",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 564.507,
      "end": 566.167,
      "text": "就是它用非常小的改动",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 566.167,
      "end": 569.607,
      "text": "然后实现了对常文的一个效果非常大的提升",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 569.607,
      "end": 572.047,
      "text": "我们最后来总结一下这篇paper",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 572.047,
      "end": 574.007,
      "text": "就是长音频的识别",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 574.007,
      "end": 576.527,
      "text": "Whisper的话会有很多严重的错误",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 576.527,
      "end": 579.987,
      "text": "但这些错误它大概率不是模型不会识别",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 579.987,
      "end": 581.647,
      "text": "而是解码的过程中",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 581.647,
      "end": 584.847,
      "text": "被之前的识别结果和一些上下文绑架了",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.10978167989979619,
      "no_speech_prob": 1.663220776382257e-11,
      "compression_ratio": 1.3700305810397553
    },
    {
      "start": 584.847,
      "end": 587.127,
      "text": "导致它越识别越错",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08998628493842728,
      "no_speech_prob": 1.183921970876467e-11,
      "compression_ratio": 1.2037037037037037
    },
    {
      "start": 587.127,
      "end": 587.947,
      "text": "越识别越错",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08998628493842728,
      "no_speech_prob": 1.183921970876467e-11,
      "compression_ratio": 1.2037037037037037
    },
    {
      "start": 587.947,
      "end": 591.127,
      "text": "所以作者他不是去强化原始模型的能力",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08998628493842728,
      "no_speech_prob": 1.183921970876467e-11,
      "compression_ratio": 1.2037037037037037
    },
    {
      "start": 591.127,
      "end": 592.627,
      "text": "也就是没有训练原始的模型",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08998628493842728,
      "no_speech_prob": 1.183921970876467e-11,
      "compression_ratio": 1.2037037037037037
    },
    {
      "start": 592.627,
      "end": 596.827,
      "text": "而是直接在token选择的阶段做干预",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08998628493842728,
      "no_speech_prob": 1.183921970876467e-11,
      "compression_ratio": 1.2037037037037037
    },
    {
      "start": 596.827,
      "end": 599.267,
      "text": "这个角度就是在工程测作",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08998628493842728,
      "no_speech_prob": 1.183921970876467e-11,
      "compression_ratio": 1.2037037037037037
    },
    {
      "start": 599.267,
      "end": 600.387,
      "text": "然后也很聪明",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08998628493842728,
      "no_speech_prob": 1.183921970876467e-11,
      "compression_ratio": 1.2037037037037037
    },
    {
      "start": 600.387,
      "end": 603.007,
      "text": "好这期视频就给大家介绍到这",
      "chunk": 1,
      "language": "zh",
      "avg_logprob": -0.08998628493842728,
      "no_speech_prob": 1.183921970876467e-11,
      "compression_ratio": 1.2037037037037037
    }
  ],
  "segmented_transcription": {
    "status": "SEGMENTED_TRANSCRIPT_OK",
    "created_at": "2026-09-06T17:57:39",
    "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260906_175655__不用训练模型_Whisper性能暴涨_背后只改了一步解码/segmented_work/chunk_manifest.json",
    "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260906_175655__不用训练模型_Whisper性能暴涨_背后只改了一步解码/audio_16k_mono.wav",
    "audio_sha256": "7c440e1e41b9c2963e71143ffc98c724f5b8fdfa692cb3b8e4c4b97e6339715e",
    "audio_duration": 605.947938,
    "model": "/Users/sagawa/AI_WORK/cache/huggingface/hub/models--mlx-community--whisper-large-v3-turbo/snapshots/a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb",
    "language_hint": "zh",
    "condition_on_previous_text": false,
    "chunk_count": 2,
    "resumed_chunks": 0,
    "merged_segment_count": 279,
    "cleaned_segment_count": 279,
    "removed_hallucination_segment_count": 0,
    "final_segment_end": 603.007,
    "global_repeat_run": 1,
    "global_repeat_text": "大家好,今天给大家介绍一篇比较新的paper,也是基于Visper作语音识别的。这篇论文的题目叫Visper CD,专门解决长音频的语音识别问题。如果用一句话来总结的话,这篇paper它的核心亮点就在于不改变Visper模型的参数,也无需训练。只在Visper解码时给它加一个反向的揪片器,这样专门来压制长音频的语音识别。",
    "tail_repeat_run": 1,
    "tail_repeat_text": "从500多变成57",
    "hallucination_region_count": 0,
    "hallucination_regions": [],
    "qc_ok": true,
    "qc_errors": []
  }
}