「悉」被读成了 xì「悉」 is read xì
字幕是「悉」,配音把一声念成了四声。只重写这一个音节,前后的「已」和「数」仍是原声。The subtitle says 悉 (xī); the voice gives it the fourth tone. Only that syllable is rewritten — 已 and 数 on either side stay original.
AI 读错一个字,不必重做整个镜头。VoxPatch 只重写错音的那一小段,保留原始音色、语气和背景声,画面一个像素都不动。When AI misreads one word, you should not have to regenerate the whole shot. VoxPatch rewrites only the faulty syllables, keeping the original timbre, delivery and background. Not a single pixel moves.
每个案例都是同一段原片:切换「原音 / 修正后」,画面对照,声音对比。Every case uses the same source clip: switch between original and repaired and compare the sound, with the picture held constant.
字幕是「悉」,配音把一声念成了四声。只重写这一个音节,前后的「已」和「数」仍是原声。The subtitle says 悉 (xī); the voice gives it the fourth tone. Only that syllable is rewritten — 已 and 数 on either side stay original.
字幕写的是「更」,配音念成了 xī。只重写这个音节的读音,句子其余部分仍是原声。The subtitle says 更 (gēng); the voice reads xī. Only that syllable is rewritten — the rest of the line stays original.
一个名字里只有一个音节坏了。只替换那一个字,前后字、连读和语气保持原声。Only one syllable in the name is broken. Just that character is replaced; the neighbours, liaison and delivery stay original.
不需要理解模型,也不需要重建时间线。工作台里只有上传、试听和采纳。No model knowledge, no timeline rebuild. The workbench is upload, audition, accept.
拖进工作台,或粘贴 YouTube 链接。画面带中文字幕的视频可以直接进入流水线。Drop a video into the workbench or paste a YouTube link. Footage with burned-in Chinese subtitles goes straight into the pipeline.
读字幕、听发音,找出读错的音节;只重写出问题的那一小段,其余原样保留。Subtitles are read and the voice is analysed; only the faulty syllable is rewritten, everything else is left untouched.
报告页逐条 A/B 试听,接受或拒绝。采纳后导出成片,视频流不重新编码。Audition every fix A/B in the report, accept or reject. Export keeps the original video stream — no re-encode.
把「整段重做」和「局部修音」放在一起看,差别不在模型大小,而在改动范围。Put a full regeneration next to a local repair: the difference is not model size, it is how much of the take you are willing to replace.
| 重新生成整镜Regenerate the shot | VoxPatch | |
|---|---|---|
| 画面Picture | 整段重新编码,人物、构图、光影都可能变。The whole shot is re-encoded; faces, framing and light can all drift. | 视频流直接复制,成片前后码流 md5 一致。The video stream is copied verbatim; the md5 matches before and after. |
| 表演Performance | 语气、节奏、停顿重新采样,等于再演一遍。Delivery, rhythm and pauses are resampled — the take is performed again. | 原表演保留,只替换出问题的那几个音节。The original performance stays; only the faulty syllables are replaced. |
| 背景Background | 音乐、环境声一起重做,口型与配乐要重新对。Music and ambience are rebuilt, and sync has to be redone. | 源分离后只动人声,伴奏与画内声留在原位。Source separation touches the vocal only; music and diegetic sound stay put. |
| 审核Review | 整条重看一遍,才知道这次有没有新的错。The whole shot has to be watched again to catch new mistakes. | 逐条试听,接受或拒绝,改动范围一眼可见。Every fix is auditioned and accepted or rejected — the blast radius is visible. |
| 返工Rollback | 每次重生成都是新的不确定性。Every regeneration is a fresh roll of the dice. | 自检不过就保留原音,不会把不可懂的合成音剪进去。If self-check fails, the original audio is kept — unintelligible synthetic speech is never spliced in. |
以上是流水线的设计约束,不是第三方对比数据;每条都对应报告里可核对的项目。These are design constraints of the pipeline, not third-party benchmark numbers; each one maps to a check visible in the report.
不改变创作流程,只在交付前多一道可控的修音步骤。It does not change the creative workflow — it adds one controllable repair step before delivery.
一个字的读音不该毁掉一条过了的镜头。修完直接进剪辑。One misread word should not cost a good take. Fix it and keep cutting.
角色声线固定,局部修音比整句重配更容易保持一致。With a fixed character voice, a local repair stays closer to the original than re-recording the line.
多条片子排队处理,每条都能单独审核、单独导出。Queue several videos, then review and export each one on its own.
同一套流水线可以接进现有的生产工具;对外 API 尚未发布。The same pipeline can sit behind existing production tools; the public API is not released yet.
内测期间不收费;账号与个人工作区还在准备中,暂未开放自助注册。正式版本会按修复点或时长计费;团队与 API 方案尚未发布。The beta is free; accounts and personal workspaces are being prepared, so self sign-up is not open yet. The released version will charge per repair or per minute; team and API plans are not published yet.
内测期间不收费;账号与个人工作区准备中。Free during the beta; accounts and personal workspaces are being prepared.
正式版拟按修复点或视频时长计费。Planned to charge per repair or per minute of video.
批量队列、成员审核与接口集成。Batch queues, member review and API integration.
页面只展示当前真实存在的方案;未发布的收费方式不会写成价格。Only plans that actually exist today are shown. Unreleased pricing is not invented here.
下面每一条都对应工作台和报告里的真实行为。Each answer below matches what the workbench and report actually do.
画面里必须带中文字幕:流水线用 OCR 读画面字幕作为参考文本。支持 .mp4 .mov .mkv .webm .m4v .ts。The footage must have burned-in Chinese subtitles — the pipeline reads them with OCR as reference text. Supported: .mp4 .mov .mkv .webm .m4v .ts.
不会。成片用 -c:v copy 封装,视频基本流前后做 md5 比对,报告里会给出「视频码流未变」。No. The output is muxed with -c:v copy; the video elementary stream is md5-compared before and after, and the report states that it is unchanged.
补丁必须通过重识别、声调与声纹自检才会落地;不过就保留原音,并在报告里标成待复核。不会把听不懂的合成音剪进成片。A patch lands only after re-recognition, tone and speaker self-checks. Otherwise the original audio is kept and the item is marked for review — unintelligible synthetic speech is never spliced in.
内测期间全部免费,采纳与否都不计费。正式计费规则发布前会单独说明。The beta is free; accepting or rejecting a fix costs nothing. Formal billing rules will be announced before they take effect.
只落在本机的 jobs/uploads/ 目录,工作台可以逐条删除。模型在本机或本机容器内运行,素材不会上传到外部推理服务。Only under jobs/uploads/ on this machine, and each record can be deleted from the workbench. Models run locally or in local containers; footage is not sent to an external inference service.
请只上传你拥有权利或已获授权的素材。工具只改音轨,不改变素材的版权归属。Only upload material you own or are licensed to use. The tool changes the audio track; it does not grant any rights to the footage.
Fix the voice. Keep the scene.
内测开放后,上传一段带中文字幕的片子,听一遍前后对比,再决定采不采纳。Once the beta opens, upload a clip with Chinese subtitles, audition the difference, then decide.