[{"data":1,"prerenderedAt":295},["ShallowReactive",2],{"guide-zh-Hans-clean-up-a-field-recording":3,"guide-zh-Hans-clean-up-a-field-recording-related":285},{"id":4,"title":5,"author":6,"body":7,"date":265,"description":266,"difficulty":267,"extension":268,"image":269,"lastModified":270,"meta":271,"navigation":272,"outcome":273,"path":274,"readingTime":275,"robots":269,"scenario":276,"seo":277,"stem":279,"tools":280,"__hash__":284},"guides_zh_Hans\u002Fzh-Hans\u002F8.guides\u002F7.clean-up-a-field-recording.md","录音降噪教程：先降噪还是先提音量，顺序错了全白做","FlowPick Team",{"type":8,"value":9,"toc":254},"minimark",[10,14,17,21,24,31,37,43,49,52,56,59,68,75,81,84,88,91,97,102,105,117,120,126,132,138,143,146,150,158,161,171,177,182,185,189,192,203,209,215,219,225,231,237,248,251],[11,12,13],"p",{},"真房间里录出来的声音，和你想象的不一样。有风扇、有冰箱、有车流，说话的人离麦克风三米远。通常两个问题一起来：太小声，也太吵。",[11,15,16],{},"直觉是先调大音量，因为那是最明显的问题。这个顺序是错的，后面会解释为什么。",[18,19,20],"h2",{"id":20},"降噪能做和不能做的",[11,22,23],{},"降噪跑的是 RNNoise，按短帧处理，判断哪部分是人声、哪部分是噪声。这个设计直接决定了它的边界：",[11,25,26,30],{},[27,28,29],"strong",{},"稳态噪声处理得好。"," 风扇声、空调声、房间底噪、远处的车流——频谱稳定的都行。",[11,32,33,36],{},[27,34,35],{},"背景里的说话声处理得差。"," 那是语音不是噪声，模型会保留下来，而且可能把人声弄坏。",[11,38,39,42],{},[27,40,41],{},"被盖住的字救不回来。"," 噪声把某个音节完全盖住，那个音节就是没了，任何处理都还原不出。",[11,44,45,48],{},[27,46,47],{},"会损伤音乐。"," 录音里有音乐的话，处理后通常会变怪。这是个人声工具。",[11,50,51],{},"先判断你手上是哪种情况，再决定要不要走这套流程。",[18,53,55],{"id":54},"第一步把音频从视频里取出来","第一步：把音频从视频里取出来",[11,57,58],{},"现场录音常常装在视频容器里：手机拍的视频、录屏、相机文件。",[11,60,61,62,67],{},"用",[63,64,66],"a",{"href":65},"\u002Fzh-Hans\u002Ftools\u002Faudio-extractor","音频提取","把音轨取出来。后面每一步都是跟一个 40 MB 的音频打交道，而不是 2 GB 的视频，速度和整理难度都不一样。",[11,69,70],{},[71,72],"img",{"alt":73,"src":74},"提取后的 M4A 音频结果，显示 AAC 编码、原始质量、无需重新编码和 33.1 KB 体积","\u002Fimg\u002Fguides\u002Faudio-extractor-output.jpg",[11,76,77],{},[78,79,80],"em",{},"英文界面，生成的四秒 MP4 → 33.1 KB AAC\u002FM4A 音频。Original audio（原始音频）模式直接拷贝音频数据，不重新编码。",[11,82,83],{},"源本来就是音频文件的话，跳过这步。",[18,85,87],{"id":86},"第二步先降噪再提音量","第二步：先降噪，再提音量",[11,89,90],{},"这是决定结果的一步。",[11,92,93],{},[71,94],{"alt":95,"src":96},"两条路径对比：先放大音量会把底噪一起放大；先降噪再放大，结果更干净","\u002Fimg\u002Fguides\u002Fdenoise-before-gain.svg",[11,98,99],{},[78,100,101],{},"先加增益，噪声也一起加。降噪器面对的是一个更响但同样脏的信号，判断并不会更准。",[11,103,104],{},"先调大音量，所有东西都跟着变大，噪声也一样。降噪器随后看到的是一个更响但噪声比例没变的信号，它的估计不会更好——而你的耳朵这时候更容易听出处理痕迹。",[11,106,107,108,112,113,116],{},"所以顺序是：用",[63,109,111],{"href":110},"\u002Fzh-Hans\u002Ftools\u002Faudio-noise-remover","降噪","打开",[27,114,115],{},"原始电平","的文件处理，干净了之后再谈音量。",[11,118,119],{},"两个设置要留意：",[11,121,122,125],{},[27,123,124],{},"先预览。"," 工具默认处理前 30 秒作为预览。长文件一定要先听预览，因为全量处理要等很久，而结果未必更好。",[11,127,128,131],{},[27,129,130],{},"降噪比例。"," 输出是按百分比混合的，值越低，混回原音的部分越多。拉满降得最干净，但人声的质感也去得最多，听起来像被处理过。准备发出去的素材，一般不要拉满；只准备用来转录的，可以拉高一些。",[11,133,134],{},[71,135],{"alt":136,"src":137},"RNNoise 界面，显示 80% 降噪混合比例、原音与降噪试听，以及 WAV 导出","\u002Fimg\u002Fguides\u002Fnoise-remover-compare.jpg",[11,139,140],{},[78,141,142],{},"英文界面，使用生成语音、80% 混合比例演示。这不是嘈杂的现场录音，也不是语音质量改善的实测证据。",[11,144,145],{},"工具单文件上限 30 分钟。更长的要切分——顺便也是把用不上的部分去掉的理由。",[18,147,149],{"id":148},"第三步小幅度提音量","第三步：小幅度提音量",[11,151,152,153,157],{},"现在录音干净了，但还是小声。用",[63,154,156],{"href":155},"\u002Fzh-Hans\u002Ftools\u002Faudio-volume-booster","音量提升","处理。",[11,159,160],{},"提到听着舒服就行，不要提到响。两个原因：",[162,163,164,168],"ul",{},[165,166,167],"li",{},"提音量也会把残留的噪声一起提上来，而降噪之后残留的往往是人声换气那种嘶声，模型没能分离干净。",[165,169,170],{},"一旦削顶——峰值撞到上限——损伤是永久的。留出余量。",[11,172,173],{},[71,174],{"alt":175,"src":176},"音量调整界面，显示增加三分贝、峰值保护，以及原音与处理后音频试听","\u002Fimg\u002Fguides\u002Fvolume-booster-gain.jpg",[11,178,179],{},[78,180,181],{},"英文界面，生成语音使用 +3 dB 增益并开启峰值保护。显示的 61% 是源音频峰值分析；当前界面没有显示实测输出峰值。",[11,183,184],{},"注意齿音：\"s\"和\"sh\"这类音是最先失真的，增益给多了就刺耳。",[18,186,188],{"id":187},"第四步先想清楚这个文件要干什么用","第四步：先想清楚这个文件要干什么用",[11,190,191],{},"同一段录音，用途不同，处理方式就不同。导出之前先明确：",[11,193,194,197,198,202],{},[27,195,196],{},"只用来转录。"," 降噪可以重一些，音量适中就好。识别对底噪的敏感度高于绝对响度。处理完交给",[63,199,201],{"href":200},"\u002Fzh-Hans\u002Ftools\u002Faudio-to-text","音频转文字","。",[11,204,205,208],{},[27,206,207],{},"要发出去给人听。"," 降噪轻一些，音量做统一。过度处理的人声，听感比轻微的背景噪声更糟。",[11,210,211,214],{},[27,212,213],{},"还要继续在剪辑软件里用。"," 导出处理后的 WAV，最后那几 dB 留给剪辑软件做。别在这里一步到位。",[18,216,218],{"id":217},"处理结果不对劲先对号入座","处理结果不对劲？先对号入座",[11,220,221,224],{},[27,222,223],{},"听起来像机器人。"," 降噪过头了。回到原文件、降低比例重跑；这个处理不是增量的，必须从源头重来。",[11,226,227,230],{},[27,228,229],{},"噪声还在。"," 多半是它本来就不是稳态噪声，或者相对人声太响了。判断一下是风扇声（该有效果）还是旁边有人说话（不会有）。",[11,232,233,236],{},[27,234,235],{},"人声又远又薄。"," 麦克风离得远。降噪加不了距离感，它只能去掉不是人声的部分，而远处的说话本来就没多少\"贴脸\"的成分。",[11,238,239,242,243,247],{},[27,240,241],{},"文件变大了。"," 降噪输出固定是 WAV，不管输入是什么格式。在意体积就用",[63,244,246],{"href":245},"\u002Fzh-Hans\u002Ftools\u002Faudio-converter","音频转换","转成 MP3 或 M4A。",[18,249,250],{"id":250},"期望值放在哪",[11,252,253],{},"一段对它的用途来说够用的录音，以及一个诚实的判断：隔着房间用手机录的，不会变成演播室访谈，但它可以变成一份完全可搜索的文稿——而那往往才是真正的目的。",{"title":255,"searchDepth":256,"depth":256,"links":257},"",2,[258,259,260,261,262,263,264],{"id":20,"depth":256,"text":20},{"id":54,"depth":256,"text":55},{"id":86,"depth":256,"text":87},{"id":148,"depth":256,"text":149},{"id":187,"depth":256,"text":188},{"id":217,"depth":256,"text":218},{"id":250,"depth":256,"text":250},"2026-10-29","手机在现场录的音又小又吵。稳态噪声降得掉、人声干扰降不掉；降噪必须在提音量之前，这个顺序比任何参数都重要。从提取音轨、降噪、增益到按用途导出的完整流程。","basic","md",null,"2026-11-01",{},true,"一段音量和干净程度都够用、能转录也能发出去的语音录音。","\u002Fzh-hans\u002Fguides\u002Fclean-up-a-field-recording",6,"edit",{"description":278,"title":5},"手机现场录音降噪和音量提升教程：先用音频提取把音轨从视频里取出来，再用本地 RNNoise 降噪（对风扇、空调、底噪等稳态噪声有效，对背景人声无效），最后小幅度提音量避免削顶失真。包含降噪比例怎么调、为什么必须先降噪再增益、按转录或发布选择处理强度，以及处理后像机器人等问题的处理。","zh-Hans\u002F8.guides\u002F7.clean-up-a-field-recording",[281,282,283],"\u002Ftools\u002Faudio-extractor","\u002Ftools\u002Faudio-noise-remover","\u002Ftools\u002Faudio-volume-booster","vKp0voC28WD_c1W4K1dUsGdZ-XQbHA6UJHHvlAS-4Lc",[286,293],{"path":287,"title":288,"description":289,"tools":290},"\u002Fzh-hans\u002Fguides\u002Fseparate-vocals-from-a-track","人声分离怎么做？在浏览器里把歌曲拆成人声和伴奏","用本地跑的 AI 模型把混音拆成人声、伴奏两条 WAV：单次十分钟上限怎么绕、30 秒预览该听什么、为什么输出是两条 WAV，以及这种分离能做到和做不到的分界线。",[291,292],"\u002Ftools\u002Faudio-trimmer","\u002Ftools\u002Fvocal-separator",{"path":274,"title":5,"description":266,"tools":294},[281,282,283],1793546746476]