{
  "text": "开会更讨厌什么?除了开会本身,大概是会后整理录音。一段一个小时的多人讨论,导出录音之后,你要把它变成会议纪要。\n传统的语音转写工具确实能帮你把话变成字,但马上你就会撞上一个更头疼的问题,谁在什么时候说了什么?\n张三里四四个人轮流发言\n还有插话强话\n转写出来的文本是一锅粥\n根本分不清哪句是谁说的\n这就是所谓说话人分离难题\n最近复旦的团队放出了一个开源模型\n叫Mouse Transcribe Diaries\n他把转写和分离这两件事捏成了一个模型\n一次推理全部搞定\n而且准确率把豆包\nGPT\n4O\nGemini\n都甩在后面\n这期就来拆一拆它\n先把问题说清楚\n语音转写这件事\n技术上叫ASR\n自动语音识别\n单独看其实已经很成熟了\n你手机输入法的语音转文字就是\n但真实场景远比一个人对着话筒念要复杂\n会议、播客、访谈、电话、讲座\n这些录音有三个共同特征\n第一、长\n动不动几十分钟到一个小时\n第二,人多,两个到十几个人不等,还互相打断\n第三,杂,有背景噪音,有关门声笑声,还有大段沉默的停顿\n对这种录音,光把话转成字远远不够\n你还想知道每句话是谁说的,发生在哪一秒\n后者,有个专门的名字,叫Dialization,说话人分离\n传统做法怎么解决这个问题的?\n拼流水线\n第一步\n用VAD\n也就是语音活动检测\n把录音切成一小段一小段有人声的片段\n第二步\n用ASR模型\n把每段转成文字\n第三步\n再上独立的说话人分离模型\n给每段贴上说话人标签\n第四步\n做时间戳对齐\n把文字\n说话人\n时间这三样东西拼回去\n听起来挺合理\n问题在于\n每一步都是独立的模型\n每一步都会犯错\n而且错误会一路累积\nVAD切错了\n后面全跟着错\nASR听差了一个词\n分离模型也跟着乱\n更要命的是\n这些模型各自为正\n谁也看不到全局\n比如同一个人的两段画被切到了不同片段\n分离模型很可能把它们当成两个人\n这就是模块化方案的死学\n局部每一步都最优\n拼起来全局崩盘\n这次放模型的团队叫OpenMOSS\n中文名模斯智能\n是从复旦大学邱熙鹏教授团队孵化出来的\n邱熙鹏这个名字你可能不熟\n但他写的那本神经网络与深度学习\n是国内很多高校的教材\n他带的团队之前做过一个很有名的对话大模型\n也叫MOS\n号称中国第一个LayChatGPT的开源模型\n所以这支队伍在开源圈是有传统的\n这次他们做的Mouse Transcribe Diaries\n参数量0.9B\n大约9亿完全开源\n权重放在Hugging Face上\n协议是Apache 2.0\n意思是你可以拿去商用\n不用交一分钱\n这个体量加上这个开放程度\n对个人开发者和中小团队特别友好\n那它凭什么能把转写和分离捏在一起呢\n看架构\n它其实是把两个现成的墙模型做了架接\n音频那一侧\n用的是Whisper Medium的编码器\n就是OpenAI那个Whisper的中等版本\n负责把声音变成声学特征\n文本那一侧\n用的是QAM3 0.6B的解码器\n通一千问的小模型\n负责生成文字\n中间用一个适配器把两者连起来\n把音频特征做四倍的时间压缩\n再塞进文本模型里去\n这样做有两个直接的好处\n第一\n它能一口气吃下超长音频\n上下文窗口12万8千\n换算下来\n大约能一次处理90分钟的录音\n完全不用切片\n第二\n转写和分离是在同一个模型里联合训练的\n模型天生就知道这段话是谁的\n根本不存在事后拼接的问题\n输入是16000赫兹的单声道音频\n输出直接就是带说话人标签和时间戳的文本\n输出的格式非常紧凑\n一眼就能看懂\n它把整段录音编码成一个时间流\n每句话前面标两样东西\n一个是方括号里的时间戳\n一个是方括号里的说话人编号\n举个例子\n开头几秒\n方括号0.48\n方括号S01\n意思是0.48秒开始\n说话人1号发言\n内容是欢迎各位\n到1.66秒结束\n紧接着跳到12.26秒\n方括号S02\n说话人2号接过话茬\n这个格式可以直接为给后处理脚本\n解析成结构化的段落\n每段都带起始时间\n结束时间\n说话人\n文本这四个字段\n想导出城JSON还是字幕格式都行\n对开发者很友好\n你不用再自己写证则去硬抠时间戳了\n到了最关键的环节\n它到底有多准\n团队在四个基准上做了评测\n全是真实的长音频多人场景\n横跨中文会议和英文播客电影\n核心指标是字符错误率\n越低越好\n我挑几个最能说明问题的数字\n在Azhel 4这个中文会议基准上\n它的错误率是14.84%\n作为对比\n豆包是18.18%\nGemini 3 Pro是22.75%\n也就是说\n它一个0.9B的开源模型\n把字节和谷歌的商业模型都压在了身下\n在Movies这个英文电影基准上更夸张\n它的错误率只有6.36%\n而GPT4O是14.37\n豆包是9.94\n电影对白里有背景音\n有口音\n有重叠\n是最难啃的骨头\n它照样通吃\n这里还要特别说一个指标\n叫德尔塔CP\n它衡量的是说话人分离的额外代价\n通俗讲就是\n为了分清是谁说的\n你额外多错了多少\n它的这个值只有0.99\n几乎为0\n意味着分清说话人这件事几乎不拖累转写准确率\n这在传统流水线里是想都不敢想的\n那它跟现在最流行的开源方案比到底差在哪\n先说Whisper,OpenAI的老牌选手,能力很强\n但他只管转写\n不管分离,你想分清谁说的,得自己再外挂一个模型\n再说Whisper X,这个是社区里用的最多的组合方案\n他在Whisper基础上接了VAD和PiNote的分离模型\n确实能做对齐和分离\n但本质上还是拼接流水线\n前面说的错误累积问题\n他一个都没躲掉\n最后说PiNote\n他是专门的分离工具\n做得相当好\n但他只管分离\n转写的另找别人\n你看出规律了吗\n传统的整个生态是被切碎的\n转写归转写\n分离归分离\n你像个搬砖工一样\n把它们拼起来\n每块砖之间都有缝隙\nMoth这套做法的价值就在于\n他把这些砖直接烧成了一整块瓷砖\n从根上消除了缝隙\n这么能打\n部署难不难\n他给了三条路\n第一条\n研究用的\n直接拿原生Transformers库加载\n几行Pytho\nN代码就能跑\n适合你想看内部细节\n做二次开发的场景\n第二条\n生产环境要高吞吐\n用SGLine Omni部署\n这是一个专门的服务框架\n单张H100显卡就能跑起来\n并发16路的时候\n每秒能处理82秒的音频\n换算成时时率是0.092\n意味着处理速度是音频播放速度的10倍以上\n如果走长音频单路\n时时率最低能压到0.0197\n也就是处理速度是音频时长的50倍\n一个小时的录音一分多钟就处理完了\n第三条\n用VLLM\n适合你的环境是Cuda12\n因为SGLANG OMNI目前要求Cuda13\nVLLM的兼容性更好\n三条路\n从研究到生产\n基本把各种需求都覆盖了\n但不是所有人都愿意敲命令型\n所以团队还做了一个字幕网页应用\n叫MPD Subtitle Web\n它就是个浏览器界面\n你上传一段音频或视频\n它自动跑模型\n跑完你能在网页上直接审阅转写结果\n谁说了什么一目了然\n审完点一下导出\n格式随你挑\nJSON给程序用\nSRT和S是标准字幕格式\n直接位给播放器就行\n甚至还能用Fpack把字幕稍录进视频\n一键导出一个带字幕的MP4\n界面支持中文和英文\n会自动跟着你浏览器的语言走\n还有个很实用的小功能叫热词\n如果你录音里有专有名词\n人名\n行业术语\n模型老式识别错\n你可以在提示词里加几个热词\n引导它往正确的方向认\n这对行业内部的录音特别管用\n这种能力到底能用在哪\n我自己最看好的几个场景\n第一会议纪要\n一个小时的多人会议\n扔进去\n几分钟出来一份带说话人和时间戳的全文\n秘书或者你自己再做摘要就行\n省下来的时间是实打实的\n第二播客和视频字幕\n做内容的人最痛苦的就是给节目上字幕\n这个工具能直接出SRT\n连谁说的都给你雕好了\n第三\n访谈整理\n记者\n研究员做智性访谈\n几十段录音要转写要编码\n这是真正能解放生产力的工具\n第四\n更专业的领域\n法庭记录\n医疗口述\n客服质检\n这些对准确率和说话人识别要求都很高的场景\n他的杀手检就是那90分钟的长上下文\n意味着绝大部分真实录音\n它能一口气吃完\n不用你手动切块\n想试的话\n装起来不复杂\n前提是Python 3.12\n用UV这个包管理器\n建一个虚拟环境\n然后从GitHub客龙仓库\n一行UVPIP Install装好依赖\n它会自动帮你处理PyTorch的GPU后端\n装完有两条路\n想看代码细节的\n直接用Transformers加在模型跑推里\n想跑成服务的\n用SGLang Omni或者VLLM启动一个API服务\n调用方式兼容OpenAI那套音频接口\n你原来用Whisper API的代码\n改个地址就能直接迁移过来\n模型权重从Hugging Face拉\n第一次跑会下载\n大约两个G\n最后说一个让这个项目分量很重的头衔\n这个模型7月9号才开源\n到现在不到20天\nGitHub已经冲到1300多star\n还登上了Trendshift当日Python仓库的第11名\n更硬的成绩是\n2026年语音领域的鼎会Interspeech\n办了第二届多语言低资源语音挑战赛\n他拿了第一名覆盖14种语言\n一个久易参数的开源模型\n在长音频多说话人这个最难的真实场景上\n同时打赢了准确率和效率两场仗\n还把代码和权重都敞开了\n代码在GitHub\n搜OpenMOSS\n斜杠\nMouseTranscribedDiaries就能找到\n权重在HuggingFace\n如果你平时要跟长录音打交道\n做字幕\n做记要\n做访谈整理\n这个工具真的值得装一个试试\n这期到这里\n我们下期见",
  "language": "multilingual",
  "duration": 801.024,
  "segments": [
    {
      "start": 5.208,
      "end": 16.318,
      "text": "开会更讨厌什么?除了开会本身,大概是会后整理录音。一段一个小时的多人讨论,导出录音之后,你要把它变成会议纪要。",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 17.098,
      "end": 25.778,
      "text": "传统的语音转写工具确实能帮你把话变成字,但马上你就会撞上一个更头疼的问题,谁在什么时候说了什么?",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 25.778,
      "end": 28.598,
      "text": "张三里四四个人轮流发言",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 28.598,
      "end": 29.998,
      "text": "还有插话强话",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 29.998,
      "end": 32.358,
      "text": "转写出来的文本是一锅粥",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 32.358,
      "end": 35.338,
      "text": "根本分不清哪句是谁说的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 35.338,
      "end": 38.358,
      "text": "这就是所谓说话人分离难题",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 38.358,
      "end": 42.058,
      "text": "最近复旦的团队放出了一个开源模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 42.058,
      "end": 44.298,
      "text": "叫Mouse Transcribe Diaries",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 44.298,
      "end": 48.538,
      "text": "他把转写和分离这两件事捏成了一个模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 48.538,
      "end": 50.338,
      "text": "一次推理全部搞定",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 50.338,
      "end": 52.138,
      "text": "而且准确率把豆包",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 52.138,
      "end": 53.038,
      "text": "GPT",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 53.038,
      "end": 53.658,
      "text": "4O",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 53.658,
      "end": 54.398,
      "text": "Gemini",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 54.398,
      "end": 55.618,
      "text": "都甩在后面",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 55.778,
      "end": 57.778,
      "text": "这期就来拆一拆它",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 57.778,
      "end": 59.878,
      "text": "先把问题说清楚",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 59.878,
      "end": 61.918,
      "text": "语音转写这件事",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 61.918,
      "end": 63.838,
      "text": "技术上叫ASR",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 63.838,
      "end": 65.218,
      "text": "自动语音识别",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 65.218,
      "end": 67.638,
      "text": "单独看其实已经很成熟了",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 67.638,
      "end": 70.538,
      "text": "你手机输入法的语音转文字就是",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 70.538,
      "end": 74.778,
      "text": "但真实场景远比一个人对着话筒念要复杂",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 74.778,
      "end": 78.938,
      "text": "会议、播客、访谈、电话、讲座",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 78.938,
      "end": 81.298,
      "text": "这些录音有三个共同特征",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 81.298,
      "end": 82.858,
      "text": "第一、长",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 82.858,
      "end": 85.238,
      "text": "动不动几十分钟到一个小时",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 85.238,
      "end": 89.938,
      "text": "第二,人多,两个到十几个人不等,还互相打断",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 89.938,
      "end": 96.358,
      "text": "第三,杂,有背景噪音,有关门声笑声,还有大段沉默的停顿",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 96.358,
      "end": 99.998,
      "text": "对这种录音,光把话转成字远远不够",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 99.998,
      "end": 104.198,
      "text": "你还想知道每句话是谁说的,发生在哪一秒",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 104.198,
      "end": 109.438,
      "text": "后者,有个专门的名字,叫Dialization,说话人分离",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 109.938,
      "end": 112.618,
      "text": "传统做法怎么解决这个问题的?",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 113.238,
      "end": 114.138,
      "text": "拼流水线",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 114.138,
      "end": 115.358,
      "text": "第一步",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 115.358,
      "end": 116.338,
      "text": "用VAD",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 116.338,
      "end": 118.318,
      "text": "也就是语音活动检测",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 118.318,
      "end": 121.718,
      "text": "把录音切成一小段一小段有人声的片段",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 121.718,
      "end": 122.898,
      "text": "第二步",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 122.898,
      "end": 124.518,
      "text": "用ASR模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 124.518,
      "end": 126.078,
      "text": "把每段转成文字",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 126.078,
      "end": 127.518,
      "text": "第三步",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 127.518,
      "end": 130.158,
      "text": "再上独立的说话人分离模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 130.158,
      "end": 132.298,
      "text": "给每段贴上说话人标签",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 132.298,
      "end": 133.598,
      "text": "第四步",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 133.598,
      "end": 135.038,
      "text": "做时间戳对齐",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 135.038,
      "end": 135.998,
      "text": "把文字",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 135.998,
      "end": 136.918,
      "text": "说话人",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 136.918,
      "end": 138.878,
      "text": "时间这三样东西拼回去",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 138.878,
      "end": 140.738,
      "text": "听起来挺合理",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 140.738,
      "end": 141.818,
      "text": "问题在于",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 141.818,
      "end": 143.838,
      "text": "每一步都是独立的模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 143.838,
      "end": 145.358,
      "text": "每一步都会犯错",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 145.358,
      "end": 147.458,
      "text": "而且错误会一路累积",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 147.458,
      "end": 149.318,
      "text": "VAD切错了",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 149.318,
      "end": 150.738,
      "text": "后面全跟着错",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 150.738,
      "end": 153.098,
      "text": "ASR听差了一个词",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 153.098,
      "end": 154.938,
      "text": "分离模型也跟着乱",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 154.938,
      "end": 156.498,
      "text": "更要命的是",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 156.498,
      "end": 158.418,
      "text": "这些模型各自为正",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 158.418,
      "end": 159.998,
      "text": "谁也看不到全局",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 159.998,
      "end": 164.218,
      "text": "比如同一个人的两段画被切到了不同片段",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 164.218,
      "end": 167.498,
      "text": "分离模型很可能把它们当成两个人",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 167.498,
      "end": 170.158,
      "text": "这就是模块化方案的死学",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 170.158,
      "end": 171.898,
      "text": "局部每一步都最优",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 171.898,
      "end": 173.618,
      "text": "拼起来全局崩盘",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 173.618,
      "end": 177.378,
      "text": "这次放模型的团队叫OpenMOSS",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 177.378,
      "end": 179.418,
      "text": "中文名模斯智能",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 179.418,
      "end": 183.358,
      "text": "是从复旦大学邱熙鹏教授团队孵化出来的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 183.358,
      "end": 186.338,
      "text": "邱熙鹏这个名字你可能不熟",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 186.338,
      "end": 189.718,
      "text": "但他写的那本神经网络与深度学习",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 189.718,
      "end": 191.878,
      "text": "是国内很多高校的教材",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 191.878,
      "end": 195.938,
      "text": "他带的团队之前做过一个很有名的对话大模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 195.938,
      "end": 196.938,
      "text": "也叫MOS",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 196.938,
      "end": 201.298,
      "text": "号称中国第一个LayChatGPT的开源模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 201.298,
      "end": 205.158,
      "text": "所以这支队伍在开源圈是有传统的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 205.158,
      "end": 208.478,
      "text": "这次他们做的Mouse Transcribe Diaries",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 208.478,
      "end": 210.258,
      "text": "参数量0.9B",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 210.258,
      "end": 212.878,
      "text": "大约9亿完全开源",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 212.878,
      "end": 215.078,
      "text": "权重放在Hugging Face上",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 215.078,
      "end": 217.298,
      "text": "协议是Apache 2.0",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 217.298,
      "end": 219.538,
      "text": "意思是你可以拿去商用",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 219.538,
      "end": 220.898,
      "text": "不用交一分钱",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 220.898,
      "end": 224.058,
      "text": "这个体量加上这个开放程度",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 224.058,
      "end": 227.478,
      "text": "对个人开发者和中小团队特别友好",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 227.478,
      "end": 231.498,
      "text": "那它凭什么能把转写和分离捏在一起呢",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 231.498,
      "end": 232.758,
      "text": "看架构",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 232.758,
      "end": 236.798,
      "text": "它其实是把两个现成的墙模型做了架接",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 236.798,
      "end": 238.298,
      "text": "音频那一侧",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 238.298,
      "end": 240.978,
      "text": "用的是Whisper Medium的编码器",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 240.978,
      "end": 244.918,
      "text": "就是OpenAI那个Whisper的中等版本",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 244.918,
      "end": 247.318,
      "text": "负责把声音变成声学特征",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 247.318,
      "end": 248.778,
      "text": "文本那一侧",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 248.778,
      "end": 251.978,
      "text": "用的是QAM3 0.6B的解码器",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 251.978,
      "end": 253.718,
      "text": "通一千问的小模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 253.718,
      "end": 255.758,
      "text": "负责生成文字",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 255.758,
      "end": 259.198,
      "text": "中间用一个适配器把两者连起来",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 259.198,
      "end": 261.938,
      "text": "把音频特征做四倍的时间压缩",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 261.938,
      "end": 264.038,
      "text": "再塞进文本模型里去",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 264.038,
      "end": 266.738,
      "text": "这样做有两个直接的好处",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 266.738,
      "end": 267.778,
      "text": "第一",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 267.778,
      "end": 270.438,
      "text": "它能一口气吃下超长音频",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 270.438,
      "end": 272.858,
      "text": "上下文窗口12万8千",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 272.858,
      "end": 273.998,
      "text": "换算下来",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 273.998,
      "end": 276.518,
      "text": "大约能一次处理90分钟的录音",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 276.518,
      "end": 277.958,
      "text": "完全不用切片",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 277.958,
      "end": 279.018,
      "text": "第二",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 279.018,
      "end": 283.018,
      "text": "转写和分离是在同一个模型里联合训练的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 283.018,
      "end": 285.838,
      "text": "模型天生就知道这段话是谁的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 285.838,
      "end": 288.238,
      "text": "根本不存在事后拼接的问题",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 288.238,
      "end": 291.758,
      "text": "输入是16000赫兹的单声道音频",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 291.758,
      "end": 295.938,
      "text": "输出直接就是带说话人标签和时间戳的文本",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 295.938,
      "end": 298.398,
      "text": "输出的格式非常紧凑",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 298.398,
      "end": 299.798,
      "text": "一眼就能看懂",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 299.798,
      "end": 303.098,
      "text": "它把整段录音编码成一个时间流",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 303.098,
      "end": 305.418,
      "text": "每句话前面标两样东西",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 305.418,
      "end": 307.598,
      "text": "一个是方括号里的时间戳",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 307.598,
      "end": 310.278,
      "text": "一个是方括号里的说话人编号",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 310.278,
      "end": 311.698,
      "text": "举个例子",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 311.698,
      "end": 312.718,
      "text": "开头几秒",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 312.718,
      "end": 314.298,
      "text": "方括号0.48",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 314.298,
      "end": 315.918,
      "text": "方括号S01",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 315.918,
      "end": 318.558,
      "text": "意思是0.48秒开始",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 318.558,
      "end": 320.278,
      "text": "说话人1号发言",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 320.278,
      "end": 321.918,
      "text": "内容是欢迎各位",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 321.918,
      "end": 323.818,
      "text": "到1.66秒结束",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 323.818,
      "end": 326.878,
      "text": "紧接着跳到12.26秒",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 326.878,
      "end": 328.338,
      "text": "方括号S02",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 328.338,
      "end": 330.438,
      "text": "说话人2号接过话茬",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 330.438,
      "end": 334.038,
      "text": "这个格式可以直接为给后处理脚本",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 334.038,
      "end": 336.018,
      "text": "解析成结构化的段落",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 336.018,
      "end": 338.218,
      "text": "每段都带起始时间",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 338.218,
      "end": 339.358,
      "text": "结束时间",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 339.358,
      "end": 340.798,
      "text": "说话人",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 340.798,
      "end": 342.258,
      "text": "文本这四个字段",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 342.258,
      "end": 346.258,
      "text": "想导出城JSON还是字幕格式都行",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 346.258,
      "end": 348.218,
      "text": "对开发者很友好",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 348.218,
      "end": 351.558,
      "text": "你不用再自己写证则去硬抠时间戳了",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 351.558,
      "end": 353.658,
      "text": "到了最关键的环节",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 353.658,
      "end": 355.058,
      "text": "它到底有多准",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 355.058,
      "end": 357.958,
      "text": "团队在四个基准上做了评测",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 357.958,
      "end": 364.858,
      "text": "全是真实的长音频多人场景",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 364.858,
      "end": 368.338,
      "text": "横跨中文会议和英文播客电影",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 368.338,
      "end": 371.098,
      "text": "核心指标是字符错误率",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 371.098,
      "end": 372.198,
      "text": "越低越好",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 372.198,
      "end": 375.158,
      "text": "我挑几个最能说明问题的数字",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 375.158,
      "end": 378.838,
      "text": "在Azhel 4这个中文会议基准上",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 378.838,
      "end": 382.158,
      "text": "它的错误率是14.84%",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 382.158,
      "end": 382.938,
      "text": "作为对比",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 382.938,
      "end": 384.818,
      "text": "豆包是18.18%",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 384.818,
      "end": 388.238,
      "text": "Gemini 3 Pro是22.75%",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 388.238,
      "end": 389.018,
      "text": "也就是说",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 389.018,
      "end": 391.558,
      "text": "它一个0.9B的开源模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 391.558,
      "end": 395.298,
      "text": "把字节和谷歌的商业模型都压在了身下",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 395.298,
      "end": 399.038,
      "text": "在Movies这个英文电影基准上更夸张",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 399.038,
      "end": 402.098,
      "text": "它的错误率只有6.36%",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 402.198,
      "end": 405.598,
      "text": "而GPT4O是14.37",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 405.598,
      "end": 407.338,
      "text": "豆包是9.94",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 407.338,
      "end": 409.998,
      "text": "电影对白里有背景音",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 409.998,
      "end": 410.898,
      "text": "有口音",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 410.898,
      "end": 411.778,
      "text": "有重叠",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 411.778,
      "end": 413.358,
      "text": "是最难啃的骨头",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 413.358,
      "end": 414.558,
      "text": "它照样通吃",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 414.558,
      "end": 417.318,
      "text": "这里还要特别说一个指标",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 417.318,
      "end": 418.678,
      "text": "叫德尔塔CP",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 418.678,
      "end": 422.098,
      "text": "它衡量的是说话人分离的额外代价",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 422.098,
      "end": 423.318,
      "text": "通俗讲就是",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 423.318,
      "end": 425.158,
      "text": "为了分清是谁说的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 425.158,
      "end": 426.778,
      "text": "你额外多错了多少",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 426.778,
      "end": 429.518,
      "text": "它的这个值只有0.99",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 429.518,
      "end": 430.718,
      "text": "几乎为0",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 430.718,
      "end": 435.138,
      "text": "意味着分清说话人这件事几乎不拖累转写准确率",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 435.138,
      "end": 438.778,
      "text": "这在传统流水线里是想都不敢想的",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 438.778,
      "end": 443.198,
      "text": "那它跟现在最流行的开源方案比到底差在哪",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 443.198,
      "end": 447.978,
      "text": "先说Whisper,OpenAI的老牌选手,能力很强",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 447.978,
      "end": 449.438,
      "text": "但他只管转写",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 449.438,
      "end": 454.498,
      "text": "不管分离,你想分清谁说的,得自己再外挂一个模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 454.858,
      "end": 459.618,
      "text": "再说Whisper X,这个是社区里用的最多的组合方案",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 459.618,
      "end": 464.858,
      "text": "他在Whisper基础上接了VAD和PiNote的分离模型",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 464.858,
      "end": 466.998,
      "text": "确实能做对齐和分离",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 466.998,
      "end": 469.498,
      "text": "但本质上还是拼接流水线",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 469.498,
      "end": 471.558,
      "text": "前面说的错误累积问题",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 471.558,
      "end": 473.038,
      "text": "他一个都没躲掉",
      "language": "zh",
      "chunk": 0,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 473.426,
      "end": 475.306,
      "text": "最后说PiNote",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 475.306,
      "end": 477.186,
      "text": "他是专门的分离工具",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 477.186,
      "end": 478.486,
      "text": "做得相当好",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 478.486,
      "end": 479.906,
      "text": "但他只管分离",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 479.906,
      "end": 481.446,
      "text": "转写的另找别人",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 481.446,
      "end": 483.186,
      "text": "你看出规律了吗",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 483.186,
      "end": 486.066,
      "text": "传统的整个生态是被切碎的",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 486.066,
      "end": 487.306,
      "text": "转写归转写",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 487.306,
      "end": 488.606,
      "text": "分离归分离",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 488.606,
      "end": 490.726,
      "text": "你像个搬砖工一样",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 490.726,
      "end": 491.946,
      "text": "把它们拼起来",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 491.946,
      "end": 493.846,
      "text": "每块砖之间都有缝隙",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 493.846,
      "end": 496.686,
      "text": "Moth这套做法的价值就在于",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 496.686,
      "end": 499.966,
      "text": "他把这些砖直接烧成了一整块瓷砖",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 499.966,
      "end": 501.826,
      "text": "从根上消除了缝隙",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 501.826,
      "end": 503.246,
      "text": "这么能打",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 503.246,
      "end": 504.406,
      "text": "部署难不难",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 504.406,
      "end": 506.126,
      "text": "他给了三条路",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 506.126,
      "end": 507.286,
      "text": "第一条",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 507.286,
      "end": 508.326,
      "text": "研究用的",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 508.326,
      "end": 511.186,
      "text": "直接拿原生Transformers库加载",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 511.186,
      "end": 512.526,
      "text": "几行Pytho",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 512.526,
      "end": 514.406,
      "text": "N代码就能跑",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 514.406,
      "end": 516.506,
      "text": "适合你想看内部细节",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 516.506,
      "end": 518.326,
      "text": "做二次开发的场景",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 518.326,
      "end": 519.526,
      "text": "第二条",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 519.526,
      "end": 521.406,
      "text": "生产环境要高吞吐",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 521.406,
      "end": 523.606,
      "text": "用SGLine Omni部署",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 523.606,
      "end": 526.206,
      "text": "这是一个专门的服务框架",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 526.206,
      "end": 529.286,
      "text": "单张H100显卡就能跑起来",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 529.286,
      "end": 530.626,
      "text": "并发16路的时候",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 530.626,
      "end": 533.066,
      "text": "每秒能处理82秒的音频",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 533.066,
      "end": 535.986,
      "text": "换算成时时率是0.092",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 535.986,
      "end": 540.246,
      "text": "意味着处理速度是音频播放速度的10倍以上",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 540.246,
      "end": 542.366,
      "text": "如果走长音频单路",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 542.366,
      "end": 545.486,
      "text": "时时率最低能压到0.0197",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 545.486,
      "end": 548.926,
      "text": "也就是处理速度是音频时长的50倍",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 548.926,
      "end": 552.206,
      "text": "一个小时的录音一分多钟就处理完了",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 552.206,
      "end": 553.706,
      "text": "第三条",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 553.706,
      "end": 555.086,
      "text": "用VLLM",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 555.086,
      "end": 557.346,
      "text": "适合你的环境是Cuda12",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 557.346,
      "end": 561.166,
      "text": "因为SGLANG OMNI目前要求Cuda13",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 561.166,
      "end": 563.806,
      "text": "VLLM的兼容性更好",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 563.806,
      "end": 565.246,
      "text": "三条路",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 565.246,
      "end": 566.606,
      "text": "从研究到生产",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 566.606,
      "end": 569.026,
      "text": "基本把各种需求都覆盖了",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 569.026,
      "end": 572.406,
      "text": "但不是所有人都愿意敲命令型",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 572.406,
      "end": 576.006,
      "text": "所以团队还做了一个字幕网页应用",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 576.006,
      "end": 578.066,
      "text": "叫MPD Subtitle Web",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 578.066,
      "end": 580.246,
      "text": "它就是个浏览器界面",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 580.246,
      "end": 582.606,
      "text": "你上传一段音频或视频",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 582.606,
      "end": 584.126,
      "text": "它自动跑模型",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 584.126,
      "end": 587.926,
      "text": "跑完你能在网页上直接审阅转写结果",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 587.926,
      "end": 590.106,
      "text": "谁说了什么一目了然",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 590.106,
      "end": 592.086,
      "text": "审完点一下导出",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 592.086,
      "end": 593.386,
      "text": "格式随你挑",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 593.386,
      "end": 595.406,
      "text": "JSON给程序用",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 595.406,
      "end": 598.466,
      "text": "SRT和S是标准字幕格式",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 598.466,
      "end": 600.546,
      "text": "直接位给播放器就行",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 600.546,
      "end": 604.946,
      "text": "甚至还能用Fpack把字幕稍录进视频",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 604.946,
      "end": 607.526,
      "text": "一键导出一个带字幕的MP4",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 607.526,
      "end": 609.966,
      "text": "界面支持中文和英文",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 609.966,
      "end": 612.626,
      "text": "会自动跟着你浏览器的语言走",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 612.626,
      "end": 615.526,
      "text": "还有个很实用的小功能叫热词",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 615.526,
      "end": 617.926,
      "text": "如果你录音里有专有名词",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 617.926,
      "end": 618.706,
      "text": "人名",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 618.706,
      "end": 619.846,
      "text": "行业术语",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 619.846,
      "end": 621.386,
      "text": "模型老式识别错",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 621.386,
      "end": 624.246,
      "text": "你可以在提示词里加几个热词",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 624.246,
      "end": 626.566,
      "text": "引导它往正确的方向认",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 626.566,
      "end": 629.846,
      "text": "这对行业内部的录音特别管用",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 629.846,
      "end": 632.626,
      "text": "这种能力到底能用在哪",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 632.626,
      "end": 635.246,
      "text": "我自己最看好的几个场景",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 635.246,
      "end": 637.326,
      "text": "第一会议纪要",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 637.326,
      "end": 639.146,
      "text": "一个小时的多人会议",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 639.146,
      "end": 640.086,
      "text": "扔进去",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 640.086,
      "end": 643.826,
      "text": "几分钟出来一份带说话人和时间戳的全文",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 643.826,
      "end": 646.706,
      "text": "秘书或者你自己再做摘要就行",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 646.706,
      "end": 649.726,
      "text": "省下来的时间是实打实的",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 649.726,
      "end": 652.446,
      "text": "第二播客和视频字幕",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 652.446,
      "end": 656.226,
      "text": "做内容的人最痛苦的就是给节目上字幕",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 656.226,
      "end": 658.706,
      "text": "这个工具能直接出SRT",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 658.706,
      "end": 660.866,
      "text": "连谁说的都给你雕好了",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 660.866,
      "end": 662.006,
      "text": "第三",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 662.006,
      "end": 663.206,
      "text": "访谈整理",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 663.206,
      "end": 663.986,
      "text": "记者",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 663.986,
      "end": 665.906,
      "text": "研究员做智性访谈",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 665.906,
      "end": 668.386,
      "text": "几十段录音要转写要编码",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 668.386,
      "end": 670.986,
      "text": "这是真正能解放生产力的工具",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 670.986,
      "end": 672.086,
      "text": "第四",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 672.086,
      "end": 673.526,
      "text": "更专业的领域",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 673.526,
      "end": 674.606,
      "text": "法庭记录",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 674.606,
      "end": 675.686,
      "text": "医疗口述",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 675.686,
      "end": 676.786,
      "text": "客服质检",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 676.786,
      "end": 681.546,
      "text": "这些对准确率和说话人识别要求都很高的场景",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 681.546,
      "end": 685.366,
      "text": "他的杀手检就是那90分钟的长上下文",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 685.366,
      "end": 687.746,
      "text": "意味着绝大部分真实录音",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 687.746,
      "end": 689.226,
      "text": "它能一口气吃完",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 689.226,
      "end": 690.846,
      "text": "不用你手动切块",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 690.846,
      "end": 692.326,
      "text": "想试的话",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 692.326,
      "end": 693.706,
      "text": "装起来不复杂",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 693.706,
      "end": 696.366,
      "text": "前提是Python 3.12",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 696.366,
      "end": 698.346,
      "text": "用UV这个包管理器",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 698.346,
      "end": 699.706,
      "text": "建一个虚拟环境",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 699.706,
      "end": 702.626,
      "text": "然后从GitHub客龙仓库",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 702.626,
      "end": 705.306,
      "text": "一行UVPIP Install装好依赖",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 705.306,
      "end": 709.006,
      "text": "它会自动帮你处理PyTorch的GPU后端",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 709.006,
      "end": 710.946,
      "text": "装完有两条路",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 710.946,
      "end": 712.986,
      "text": "想看代码细节的",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 712.986,
      "end": 716.386,
      "text": "直接用Transformers加在模型跑推里",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 716.386,
      "end": 718.186,
      "text": "想跑成服务的",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 718.186,
      "end": 723.826,
      "text": "用SGLang Omni或者VLLM启动一个API服务",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 723.826,
      "end": 727.406,
      "text": "调用方式兼容OpenAI那套音频接口",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 727.406,
      "end": 730.366,
      "text": "你原来用Whisper API的代码",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 730.366,
      "end": 732.986,
      "text": "改个地址就能直接迁移过来",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 732.986,
      "end": 735.646,
      "text": "模型权重从Hugging Face拉",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 735.646,
      "end": 737.166,
      "text": "第一次跑会下载",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 737.166,
      "end": 738.466,
      "text": "大约两个G",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 738.466,
      "end": 742.366,
      "text": "最后说一个让这个项目分量很重的头衔",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 742.366,
      "end": 745.186,
      "text": "这个模型7月9号才开源",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 745.186,
      "end": 747.046,
      "text": "到现在不到20天",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 747.046,
      "end": 749.946,
      "text": "GitHub已经冲到1300多star",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 749.946,
      "end": 754.646,
      "text": "还登上了Trendshift当日Python仓库的第11名",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 754.646,
      "end": 756.326,
      "text": "更硬的成绩是",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 756.326,
      "end": 759.806,
      "text": "2026年语音领域的鼎会Interspeech",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 759.806,
      "end": 763.626,
      "text": "办了第二届多语言低资源语音挑战赛",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 763.626,
      "end": 766.526,
      "text": "他拿了第一名覆盖14种语言",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 766.526,
      "end": 769.166,
      "text": "一个久易参数的开源模型",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 769.166,
      "end": 773.206,
      "text": "在长音频多说话人这个最难的真实场景上",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 773.206,
      "end": 776.806,
      "text": "同时打赢了准确率和效率两场仗",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 776.806,
      "end": 779.126,
      "text": "还把代码和权重都敞开了",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 779.126,
      "end": 780.726,
      "text": "代码在GitHub",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 780.726,
      "end": 782.406,
      "text": "搜OpenMOSS",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 782.406,
      "end": 783.486,
      "text": "斜杠",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 783.486,
      "end": 786.586,
      "text": "MouseTranscribedDiaries就能找到",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 786.586,
      "end": 788.146,
      "text": "权重在HuggingFace",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 788.146,
      "end": 791.226,
      "text": "如果你平时要跟长录音打交道",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 791.226,
      "end": 792.206,
      "text": "做字幕",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 792.206,
      "end": 793.126,
      "text": "做记要",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 793.126,
      "end": 794.446,
      "text": "做访谈整理",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 794.446,
      "end": 797.606,
      "text": "这个工具真的值得装一个试试",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 797.606,
      "end": 799.126,
      "text": "这期到这里",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    },
    {
      "start": 799.126,
      "end": 800.326,
      "text": "我们下期见",
      "language": "zh",
      "chunk": 1,
      "local_language_run": 0,
      "language_run": 0
    }
  ],
  "segmented_multilingual_transcription": {
    "status": "STABLE_SEGMENTED_MULTILINGUAL_OK",
    "created_at": "2026-09-30T23:43:08",
    "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260930_233232__一个_0.9B_开源语音模型_转写和说话人分离一次搞定_还把豆包_GPT-4o_甩在后面/segmented_work/chunk_manifest.json",
    "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260930_233232__一个_0.9B_开源语音模型_转写和说话人分离一次搞定_还把豆包_GPT-4o_甩在后面/audio_16k_mono.wav",
    "audio_sha256": "eeb86a8ec4c7f138375886296d3b426972f820bf129f35cde72761499680a517",
    "audio_duration": 801.024,
    "model": "/Users/sagawa/AI_WORK/cache/huggingface/hub/models--mlx-community--whisper-large-v3-turbo/snapshots/a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb",
    "language_profile": "/Users/sagawa/AI_WORK/video_notes/url/20260930_233232__一个_0.9B_开源语音模型_转写和说话人分离一次搞定_还把豆包_GPT-4o_甩在后面/multilingual_work/multilingual_language_profile.json",
    "dominant_languages": [
      "zh"
    ],
    "suppressed_languages": [
      "en"
    ],
    "chunk_count": 2,
    "completed_chunks": 2,
    "resumed_chunks": 0,
    "merged_segment_count": 327,
    "detected_languages": [
      "zh"
    ],
    "transcript_languages": [
      "zh"
    ],
    "missing_languages": [],
    "unexpected_languages": [],
    "language_sequence": [
      "zh"
    ],
    "speech_region_seconds": 795.816,
    "speech_covered_seconds": 792.21,
    "speech_coverage_ratio": 0.9954688018335897,
    "tail_repeat_run": 1,
    "tail_repeat_text": "开会更讨厌什么?除了开会本身,大概是会后整理录音。一段一个小时的多人讨论,导出录音之后,你要把它变成会议纪要。",
    "chunk_needs_review": [],
    "qc_ok": true,
    "qc_errors": [],
    "complete_review_artifact": true
  }
}