cut-silences

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Cut Silences (Pipeline Agent 1)

剪除静音（流水线 Agent 1）

First step of the automated edit. Takes a raw recording + its word-level transcript and removes only silence: dead air before the first word and after the last word, plus inter-word pauses longer than a threshold (trimmed down to a natural breath, never a hard zero-gap). It leaves the speaker's words untouched — false starts, retakes, and stutters are the cut-mistakes agent's job (Agent 2).

This agent is deterministic and transcript-driven, so its output (

*.silence-transcript.json

) feeds cleanly into the next agents and the beat-sync validator.

自动化编辑的第一步。接收原始录制内容和对应的逐词级转录文本，仅移除静音：首个词之前和最后一个词之后的无效空档，以及超过阈值的词间停顿（会将其缩短至自然呼吸的间隔，绝不会硬切成零间隔）。它不会改动说话者的词语——说错后重说、重新录制和结巴都由 cut-mistakes Agent（Agent 2）负责。

该 Agent 采用确定性、基于转录文本的处理方式，因此其输出（

*.silence-transcript.json

）可以直接供后续 Agent 和 beat-sync 校验器使用。

When to use

何时使用

"cut the silences", "trim the pauses", "remove dead air", "tighten the pacing"
As the first stage of the master edit workflow, right after transcription.

"剪掉静音"、"缩短停顿"、"移除无效空档"、"让节奏更紧凑"
作为主编辑工作流的第一阶段，紧接在转录之后执行。

Prerequisites

前置条件

A word-level transcript JSON. Either the ElevenLabs Scribe shape (

{ words: [{ text, start, end, type }], audio_duration_secs }

) or a generic

{ words: [{ text, start, end }] }

. Generate one with the workspace transcriber:

bash

node scripts/transcribe-elevenlabs.mjs path/to/raw.mp4    # -> path/to/raw.json

需要一个逐词级转录 JSON。可以是 ElevenLabs Scribe 的结构（

{ words: [{ text, start, end, type }], audio_duration_secs }

），也可以是通用结构

{ words: [{ text, start, end }] }

。使用工作区转录器生成：

bash

node scripts/transcribe-elevenlabs.mjs path/to/raw.mp4    # -> path/to/raw.json

Usage

用法

bash

undefined

bash

undefined

1) Plan only — compute the cut, write EDL + re-timed transcript (no video touched)

node .agents/skills/cut-silences/scripts/cut-silences.mjs <transcript.json>
--out-dir video-projects/<slug>/assets

2) With a video — also write the ffmpeg command (still does not render yet)

node .agents/skills/cut-silences/scripts/cut-silences.mjs <transcript.json>
--video video-projects/<slug>/assets/raw.mp4 --out-dir video-projects/<slug>/assets

3) Render the cut video (local ffmpeg, re-encode, A/V kept in sync)

add --apply to actually run ffmpeg

node .agents/skills/cut-silences/scripts/cut-silences.mjs <transcript.json>
--video video-projects/<slug>/assets/raw.mp4 --apply
--output video-projects/<slug>/assets/edited-silenced.mp4

undefined

node .agents/skills/cut-silences/scripts/cut-silences.mjs <transcript.json>
--video video-projects/<slug>/assets/raw.mp4 --apply
--output video-projects/<slug>/assets/edited-silenced.mp4

undefined

Options

选项

Flag	Default	Meaning
`--video <path>`	—	Source video; enables the ffmpeg command / render
`--out-dir <dir>`	next to transcript	Where outputs are written
`--output <path>`	`<video-stem>.silenced.mp4`	Cut-video path
`--gap <s>`	`0.55`	Minimum pause treated as trimmable silence
`--head-pad <s>`	`0.22`	Silence kept before the first word
`--tail-pad <s>`	`0.34`	Silence kept after the last word
`--apply`	off	Actually run ffmpeg to render the cut

标志	默认值	含义
`--video <path>`	—	源视频；用于启用 ffmpeg 命令或渲染
`--out-dir <dir>`	转录文本所在目录	输出文件的写入位置
`--output <path>`	`<video-stem>.silenced.mp4`	剪辑后视频的路径
`--gap <s>`	`0.55`	视为可裁剪静音的最短停顿
`--head-pad <s>`	`0.22`	首个词之前保留的静音
`--tail-pad <s>`	`0.34`	最后一个词之后保留的静音
`--apply`	关闭	实际运行 ffmpeg 渲染剪辑结果

How it decides (the silence rules)

判断规则（静音规则）

Pauses below
```
--gap
```
(0.55s) are left alone — natural speech rhythm.
For a trimmed pause, a natural breath is kept, scaled by context:
```
0.24s
```
for long pauses (≥2s),
```
0.20s
```
after a sentence ender (
```
. ! ?
```
),
```
0.14s
```
otherwise. The kept breath is biased slightly toward the end of the previous phrase.
Head/tail dead air is trimmed to
```
--head-pad
```
/
```
--tail-pad
```
.
Delete ranges are merged; keep ranges are the complement. The cut is rendered with an ffmpeg
```
trim
```
/
```
atrim
```
+
```
concat
```
filtergraph (written to a
```
*.silence-filter.txt
```
script and passed via
```
-/filter_complex
```
), so video and audio stay in sync.

低于
```
--gap
```
（0.55s）的停顿保持不变——保留自然的说话节奏。
对于被裁剪的停顿，会根据上下文保留一段自然呼吸：长停顿（≥2s）保留
```
0.24s
```
，句末标点（
```
. ! ?
```
）之后保留
```
0.20s
```
，其他情况保留
```
0.14s
```
。保留的呼吸间隔会略微偏向上一短语的末尾。
开头和结尾的无效空档分别裁剪至
```
--head-pad
```
/
```
--tail-pad
```
。
删除区间会进行合并，保留区间则为其补集。剪辑结果通过 ffmpeg 的
```
trim
```
/
```
atrim
```
+
```
concat
```
filtergraph 渲染（filtergraph 会写入
```
*.silence-filter.txt
```
脚本，并通过
```
-/filter_complex
```
传入），从而使视频和音频保持同步。

Outputs (written to

--out-dir

)

输出（写入

--out-dir

）

File	Purpose
`<stem>.silence-edl.json`	Keep/delete ranges, durations, params — the edit list
`<stem>.silence-transcript.json`	Words re-timed onto the edited timeline (feeds Agent 2 + beats)
`<stem>.silence-decisions.md`	Human-readable summary + largest pauses trimmed
`<stem>.silence-filter.txt`	The ffmpeg filtergraph (only with `--video` )
`<video-stem>.silenced.mp4`	The cut video (only with `--apply` )

The JSON summary printed to stdout includes

removed

removedPct

, range counts, and output paths — useful for the master workflow to log and chain.

文件	用途
`<stem>.silence-edl.json`	保留/删除区间、时长和参数——即编辑列表
`<stem>.silence-transcript.json`	按剪辑后时间线重新计时的词语（供 Agent 2 和 beat 流程使用）
`<stem>.silence-decisions.md`	便于阅读的摘要，以及被裁剪的最大停顿列表
`<stem>.silence-filter.txt`	ffmpeg filtergraph（仅在使用 `--video` 时生成）
`<video-stem>.silenced.mp4`	剪辑后的视频（仅在使用 `--apply` 时生成）

打印到 stdout 的 JSON 摘要包含

removed

、

removedPct

、区间数量和输出路径，便于主工作流记录日志和串联后续流程。

Tuning notes

调优说明

Talking-head YouTube default (
```
--gap 0.55
```
) removes roughly 15-20% of a typical raw take as pure silence. Lower
```
--gap
```
for a punchier, faster cut; raise it to preserve more natural breathing room.
If a cut feels too aggressive at sentence boundaries, raise the sentence-break breath, or raise
```
--gap
```
.

对于面对镜头的 YouTube 视频，默认设置（
```
--gap 0.55
```
）通常可从一条原始录制中裁掉约 15–20% 的纯静音。若希望剪辑更有力度、速度更快，可降低
```
--gap
```
；若希望保留更多自然呼吸空间，则提高该值。
如果剪辑在句子边界处显得过于激进，请提高句子分隔处保留的呼吸时长，或提高
```
--gap
```
。

Hand-off to the next agent

传递给下一个 Agent

Pass

<stem>.silence-transcript.json

(and the

silenced.mp4

if rendered) to the cut-mistakes agent. Because timestamps are already on the edited timeline, downstream beat timing and

scripts/validate-beat-sync.mjs

work without further adjustment.

将

<stem>.silence-transcript.json

（以及已渲染时的

silenced.mp4

）传递给 cut-mistakes Agent。由于时间戳已经基于剪辑后的时间线，下游 beat 时间对齐和

scripts/validate-beat-sync.mjs

无需进一步调整即可运行。

cut-silences

Original

Translation

Cut Silences (Pipeline Agent 1)

剪除静音（流水线 Agent 1）

When to use

何时使用

Prerequisites

前置条件

Usage

用法

1) Plan only — compute the cut, write EDL + re-timed transcript (no video touched)

1) Plan only — compute the cut, write EDL + re-timed transcript (no video touched)

2) With a video — also write the ffmpeg command (still does not render yet)

2) With a video — also write the ffmpeg command (still does not render yet)

3) Render the cut video (local ffmpeg, re-encode, A/V kept in sync)

3) Render the cut video (local ffmpeg, re-encode, A/V kept in sync)

add --apply to actually run ffmpeg

add --apply to actually run ffmpeg

Options

选项

How it decides (the silence rules)

判断规则（静音规则）

Outputs (written to
`--out-dir`
)

输出（写入
`--out-dir`
）

Tuning notes

调优说明

Hand-off to the next agent

传递给下一个 Agent

cut-silences

Original

Translation

Cut Silences (Pipeline Agent 1)

剪除静音（流水线 Agent 1）

When to use

何时使用

Prerequisites

前置条件

Usage

用法

1) Plan only — compute the cut, write EDL + re-timed transcript (no video touched)

1) Plan only — compute the cut, write EDL + re-timed transcript (no video touched)

2) With a video — also write the ffmpeg command (still does not render yet)

2) With a video — also write the ffmpeg command (still does not render yet)

3) Render the cut video (local ffmpeg, re-encode, A/V kept in sync)

3) Render the cut video (local ffmpeg, re-encode, A/V kept in sync)

add --apply to actually run ffmpeg

add --apply to actually run ffmpeg

Options

选项

How it decides (the silence rules)

判断规则（静音规则）

Outputs (written to --out-dir)

输出（写入 --out-dir）

Tuning notes

调优说明

Hand-off to the next agent

传递给下一个 Agent

Outputs (written to
`--out-dir`
)

输出（写入
`--out-dir`
）