[{"data":1,"prerenderedAt":289},["ShallowReactive",2],{"guide-zh-Hans-separate-vocals-from-a-track":3,"guide-zh-Hans-separate-vocals-from-a-track-related":278},{"id":4,"title":5,"author":6,"body":7,"date":259,"description":260,"difficulty":261,"extension":262,"image":263,"lastModified":264,"meta":265,"navigation":266,"outcome":267,"path":268,"readingTime":269,"robots":263,"scenario":270,"seo":271,"stem":273,"tools":274,"__hash__":277},"guides_zh_Hans\u002Fzh-Hans\u002F8.guides\u002F5.separate-vocals-from-a-track.md","人声分离怎么做？在浏览器里把歌曲拆成人声和伴奏","FlowPick Team",{"type":8,"value":9,"toc":247},"minimark",[10,14,18,21,38,41,44,48,54,60,63,67,76,83,89,92,100,106,111,115,118,121,127,133,139,145,150,154,157,160,166,171,174,210,213,217,223,229,235,241,244],[11,12,13],"p",{},"把混好的曲子拆成人声和伴奏，听起来像个魔术，直到你真的试一次。结果有用，但也明显不完美。开始之前分清这两点，能省掉一下午的失望。",[15,16,17],"h2",{"id":17},"这个工具实际在做什么",[11,19,20],{},"它跑的是 Kim Vocal 2，一个被训练来判断\"混音里哪部分属于人声、哪部分属于其他\"的模型。输出两条轨：",[22,23,24,32],"ul",{},[25,26,27,31],"li",{},[28,29,30],"strong",{},"人声","——模型判断为嗓音的部分",[25,33,34,37],{},[28,35,36],{},"伴奏","——模型判断不是嗓音的其余部分",[11,39,40],{},"两条都是 WAV。全程不上传，模型和运行时首次下载约 64 MB，之后在本地跑。",[11,42,43],{},"它不是什么：不是无损还原。工具自己的说明就写了，分离可能在人声里留下乐器、也可能改变人声本身。鼓和镲尤其容易漏进人声轨，处理痕迹重的唱段拆出来会偏薄。这不是参数没调好，是这个层级的技术就这样。",[15,45,47],{"id":46},"开始之前两个限制决定流程","开始之前：两个限制决定流程",[11,49,50,53],{},[28,51,52],{},"单文件最长十分钟。"," 超过就直接拒绝。这就是流程里必须有裁剪这一步的原因——不是讲究，是硬限制。",[11,55,56,59],{},[28,57,58],{},"一次一个文件，只支持桌面端。"," 这是桌面预览版，不是批量工具。扔一个文件夹进去是做不到的。",[11,61,62],{},"还有一个授权问题值得说清楚：处理你自己拥有或已获授权的音频。把一首商业曲子拆开，并不等于你获得了这两条分轨的使用权。",[15,64,66],{"id":65},"第一步裁到你要的那一段","第一步：裁到你要的那一段",[11,68,69,70,75],{},"用",[71,72,74],"a",{"href":73},"\u002Fzh-Hans\u002Ftools\u002Faudio-trimmer","音频裁剪","把源文件切到十分钟以内。",[11,77,78],{},[79,80],"img",{"alt":81,"src":82},"把一段长音频裁到分离工具十分钟上限以内的样子","\u002Fimg\u002Fguides\u002Faudio-trim-to-limit.svg",[11,84,85],{},[86,87,88],"em",{},"要的那段超过十分钟就切成两段分别处理。对这类用途来说，接缝通常不影响效果。",[11,90,91],{},"两个习惯：",[22,93,94,97],{},[25,95,96],{},"切在语句边界上，不要切在字中间。模型对连续的说话声或完整的乐句处理得更好，对从半个音节开始的碎片处理得更差。",[25,98,99],{},"裁好的文件留着。同一段输入重跑分离，结果会有细微差别，对比时需要一份稳定的参照。",[11,101,102],{},[79,103],{"alt":104,"src":105},"语音波形，起点为一秒、终点为四秒，选区时长三秒","\u002Fimg\u002Fguides\u002Faudio-trimmer-range.jpg",[11,107,108],{},[86,109,110],{},"英文界面，生成语音样本中选中了 0:01–0:04。这张短样例演示控件操作，不代表已测试十分钟音乐。",[15,112,114],{"id":113},"第二步先跑-30-秒预览再决定要不要全跑","第二步：先跑 30 秒预览，再决定要不要全跑",[11,116,117],{},"工具有前 30 秒的预览。用它。CPU 上跑全段可能是音频时长的好几倍，等二十分钟才发现输入不行，代价太大。",[11,119,120],{},"预览里重点听三件事：",[11,122,123,126],{},[28,124,125],{},"串音。"," 放人声轨，听有没有鼓、贝斯或者混响尾巴。编曲密的曲子容易漏，编曲稀疏的通常没事。",[11,128,129,132],{},[28,130,131],{},"变薄。"," 放伴奏轨，听人声是真的没了还是只是变小了。伴奏里残留人声是最常见的抱怨，这属于正常现象，不是故障。",[11,134,135,138],{},[28,136,137],{},"音量关系。"," 两条轨用同一个输出增益导出，彼此是平衡的。预览里觉得某条明显更响，那是原混音的比例，不是导出的问题。",[11,140,141],{},[79,142],{"alt":143,"src":144},"人声分离结果，可切换原音、人声和伴奏试听","\u002Fimg\u002Fguides\u002Fvocal-separator-preview.jpg",[11,146,147],{},[86,148,149],{},"英文界面，使用生成语音演示，并非混音歌曲。分离结果不能用于证明音乐分离质量。",[15,151,153],{"id":152},"第三步跑完整段并导出","第三步：跑完整段并导出",[11,155,156],{},"开始分离，等它结束。浏览器能用 GPU 时会明显快一些；核显机器上会退回 CPU，并且提示可能需要音频时长的好几倍时间。",[11,158,159],{},"结果存在浏览器的临时空间里。刷新页面能在十分钟内找回，但那是便利，不是保存。跑完立刻把两条 WAV 下载到本地。",[11,161,162],{},[79,163],{"alt":164,"src":165},"完整分离结果，显示人声与伴奏的两个 WAV 下载按钮及体积","\u002Fimg\u002Fguides\u002Fvocal-separator-download.jpg",[11,167,168],{},[86,169,170],{},"英文界面，实际完整处理了短生成语音样本。离开流程前应分别下载两个 WAV 文件。",[15,172,173],{"id":173},"拆出来的两条轨能干什么",[22,175,176,182,188,199],{},[25,177,178,181],{},[28,179,180],{},"练习","——跟着伴奏弹或唱",[25,183,184,187],{},[28,185,186],{},"翻唱或改编","——拿伴奏当底",[25,189,190,193,194,198],{},[28,191,192],{},"转录前处理","——把人声轨送去",[71,195,197],{"href":196},"\u002Fzh-Hans\u002Ftools\u002Faudio-to-text","音频转文字","，少了音乐干扰，识别通常更干净",[25,200,201,204,205,209],{},[28,202,203],{},"转格式","——WAV 很大，要分享就用",[71,206,208],{"href":207},"\u002Fzh-Hans\u002Ftools\u002Faudio-converter","音频转换","转成 MP3 或 M4A",[11,211,212],{},"转录这条值得单独提：它是少数\"分离之后另一个工具的效果真的变好\"的情况，而不只是产出一个你想要的文件。",[15,214,216],{"id":215},"分离效果不理想先对号入座","分离效果不理想？先对号入座",[11,218,219,222],{},[28,220,221],{},"文件被拒，说太长。"," 要么跳过了裁剪，要么选区本身还超过十分钟。限制算的是裁剪后的输出，不是源文件。",[11,224,225,228],{},[28,226,227],{},"人声听起来像泡在水里。"," 分离痕迹过重。换同一首曲子的别的段落试试，密的段落比稀疏的难处理。",[11,230,231,234],{},[28,232,233],{},"伴奏里还留着人声。"," 大部分素材都这样。这里没有\"把人声再压一压\"的选项——要从一个完整的混音母带里得到干净的伴奏，这个工具做不到。",[11,236,237,240],{},[28,238,239],{},"CPU 上一直跑不完。"," 取消比干等好；不会丢东西，已完成的部分会被丢弃、临时文件会释放。缩短片段，或者换一台有可用 GPU 的机器。",[15,242,243],{"id":243},"期望值放在哪",[11,245,246],{},"两条 WAV，够用于练习、当伴奏、给转录做预处理，但达不到录音室分轨的水准。这个界限就是\"浏览器本地的人声分离能做什么\"的诚实答案。",{"title":248,"searchDepth":249,"depth":249,"links":250},"",2,[251,252,253,254,255,256,257,258],{"id":17,"depth":249,"text":17},{"id":46,"depth":249,"text":47},{"id":65,"depth":249,"text":66},{"id":113,"depth":249,"text":114},{"id":152,"depth":249,"text":153},{"id":173,"depth":249,"text":173},{"id":215,"depth":249,"text":216},{"id":243,"depth":249,"text":243},"2026-10-29","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。","intermediate","md",null,"2026-11-01",{},true,"两条 WAV——人声和伴奏——来自一段你有权处理的音频。","\u002Fzh-hans\u002Fguides\u002Fseparate-vocals-from-a-track",6,"edit",{"description":272,"title":5},"人声分离和伴奏提取教程：在浏览器里本地运行 AI 模型，把混好的歌曲拆成人声轨和伴奏轨，输出两条 WAV，全程不上传。包含超过十分钟上限前的裁剪方法、用 30 秒预览判断分离质量、导出后转 MP3 分享，以及练习伴奏、翻唱底噪、转录预处理等用途和效果边界。","zh-Hans\u002F8.guides\u002F5.separate-vocals-from-a-track",[275,276],"\u002Ftools\u002Faudio-trimmer","\u002Ftools\u002Fvocal-separator","wrMM1POXB45bo38KqyAXIwmIX8Mul03bmg4KqGMJIEg",[279,281],{"path":268,"title":5,"description":260,"tools":280},[275,276],{"path":282,"title":283,"description":284,"tools":285},"\u002Fzh-hans\u002Fguides\u002Fclean-up-a-field-recording","录音降噪教程：先降噪还是先提音量，顺序错了全白做","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。",[286,287,288],"\u002Ftools\u002Faudio-extractor","\u002Ftools\u002Faudio-noise-remover","\u002Ftools\u002Faudio-volume-booster",1793546746289]