CutVoke Adds Conservative Multicamera Audio Sync for AI Editors
According to the developer’s report, CutVoke is an MIT-licensed open-source local video editor whose MCP interface allows Claude Code and Codex to edit the same timeline. Its multicam.sync command uses FFmpeg audio, 10-millisecond RMS energy envelopes, and normalized cross-correlation to align camera tracks. If confidence is below 0.65 or the best and second-best matches are insufficiently separated, it rejects the alignment and leaves the timeline unchanged. The report says this allows an agent to report ambiguity for manual alignment instead of applying an uncertain offset. The feature only analyzes the first 60 seconds, does not account for clock drift in long recordings, and accepts only continuous, normal-speed clips. The report also states that applied changes can be undone.
The reported workflow combines agent-operated editing with explicit uncertainty checks: uncertain multicamera matches are rejected, allowing manual review rather than a potentially misleading alignment. It also gives Claude Code and Codex access to a shared, undoable editing timeline while leaving camera-switch decisions to the user.
我在做 CutVoke,一个开源(MIT)的本地视频剪辑器,Claude Code / Codex 可以通过 MCP 和你编辑同一条时间线。这篇只讲其中一小块:多机位音频对齐,以及它为什么宁可拒绝也不瞎猜。
多机位剪辑里最无聊的那部分
做播客、访谈的剪辑者常描述同一种一周:剪本身还好,时间都花在对齐机位、切机位、拉平音量上。对齐机位是机械活,很适合交给软件,或者交给驱动剪辑器的 Agent。但这也是“自信地给错答案比不给答案更糟”的活:一个机位差 0.3 秒,看起来就像配音片,而你可能导出后才发现。
所以 CutVoke 里 multicam.sync 的规则是:音频明显一致时才对齐,不一致就不碰时间线。
它具体做了什么
multicam.sync {referenceTrackId, targetTrackId, maxOffsetSeconds?}(源码:src/cutvoke/rendering/multicam_sync.py):