remove-silences

name

remove-silences

description

Loudness-based silence removal on the live Adobe Premiere Pro timeline via the OpenCutAgent "Remove Silences" panel. Measure audio levels (ffmpeg, NO transcription/API), recommend or set the Noise Threshold ("Suggest threshold"), preview the silence/margin zones, and ripple/lift/mute-delete dead air. Use when the user wants to remove silences, cut dead air / pauses / gaps of quiet, tighten the pacing, set or auto-calculate a noise threshold, or asks about the Remove Silences tab / "Suggest threshold". Requires the OpenCutAgent panel open in Premiere and the `premiere` MCP server connected.

Remove Silences (loudness-based)

The Remove Silences tab is the first thing most users reach for: it removes dead air by audio loudness, not transcription. The server measures each clip's source loudness with ffmpeg (cached, no Scribe call, no API key, no credits), the panel draws the dB envelope, and silences below the Noise Threshold are removed — ripple, lift, or mute. This is faster and cheaper than the transcript-based ppro_remove_silences; use it for "cut the silences / dead air / pauses". Use the transcript path only when the user specifically wants filler words ("um", "uh") cut, which need the words.

Edits land in the user's real sequence, non-destructively. Undo is always Cmd+Z in Premiere.

The controls (mirror the panel)

• Noise Threshold (dB) — audio below this is silence. The one knob that matters most.
• Remove Silences Longer Than (ms) — only cut silences at least this long. Default 120.
• Keep speech longer than (ms) — speech islands shorter than this (pops, mic bumps, chatter bursts) are treated as silence, unconditionally (exact AutoCut/TimeBolt semantics — merges a stretch of sub-keepTalk chatter into one solid removable block). Values > ~750ms risk eating short connector words. Default 400.
• Margin before / Margin after (ms) — air kept before the next / after the previous speech (the green "Margins"). Default 120/120.
• Silence Management — remove (ripple, close gaps) · keep_spaces (lift, leave gaps) · mute (silence, keep the picture) · keep (no change).
• Transitions — recorded; v1 applies clean cuts (no auto-crossfade yet). Tell the user if they pick J/L-Cut etc.

Tools

• ppro_analyze_audio_levels — measure loudness; returns per-clip and overall stats (min / median / speechDb / noiseFloorDb) and a suggested threshold. This powers the panel's "Suggest threshold." Pass set_threshold_db to push YOUR chosen dB live into the panel.
• ppro_remove_silences_by_level — the full cut: preview with dry_run:true (returns the cut list + total seconds), then dry_run:false to apply. Controls: threshold_db, min_silence_ms, keep_talk_ms, margin_before_ms, margin_after_ms, mode (remove/keep_spaces/mute).

Both default clip_id:"all" (every video clip with media). They share the same loudness engine as the panel, so chat and panel always agree.

"Suggest threshold" — the flow the user expects

When the user clicks Suggest threshold in the panel, or asks you to "calculate / set the silence threshold with AI":

1. ppro_analyze_audio_levels — read noiseFloorDb and speechDb (overall and per clip).
2. Reason about it. Anchor to speech: a good threshold sits roughly 30 dB below the speech level (90th percentile), never closer than 10 dB to it, and always above the noise floor — the generous drop keeps QUIET passages (speaker leaning back, trailing off; 15–25 dB under normal level) above the threshold so they don't get shredded. The meter is peak-per-window, normalized to the recording's own peak, with a hard −60 dB floor — −60 is the slider minimum and means "detect nothing", so never answer −60; stay within −55…−20. A wide noise↔speech gap allows an aggressive (higher) threshold; a narrow gap (noisy room, music bed) needs a conservative one and a note to the user. The auto suggestedThresholdDb is a fine starting point — adjust it with judgment, don't just echo it.
3. Push your choice: call ppro_analyze_audio_levels again with set_threshold_db: <your dB> (optionally also min_silence_ms etc.). It updates the panel live — the threshold line moves and the red zones recompute. Tell the user the value and why, and that they can fine-tune and hit Remove Silences, or ask you to apply.

Headless flow (drive entirely from chat)

1. ppro_remove_silences_by_level dry_run:true — show the cut count, total seconds, and a few example timecodes. (Pass threshold_db if you've chosen one; otherwise it auto-estimates.)
2. Confirm the plan in plain English. Wait for a yes.
3. ppro_remove_silences_by_level dry_run:false with the agreed mode. Report applied count + seconds. Remind: Cmd+Z to undo.

Hard rules (non-negotiable)

1. Preview before cutting. Always dry_run:true first (or have the user review the panel's red zones), get a yes, then apply.
2. Threshold is content-specific. Never invent a dB blindly — analyze first. Re-tune from feedback ("too aggressive → lower the threshold a few dB or raise Remove Silences Longer Than"; "clipping speech → raise Margin / Keep speech longer than").
3. Respect the v1 envelope. 100%-speed clips in flat sequences. Non-100%-speed / multicam clips are skipped — say so, don't force them.
4. Transitions = clean cuts in v1. If the user selects J-Cut/L-Cut/Overlap/Constant Power, note that the style is recorded but v1 ripple/lift-deletes cleanly.
5. Undo: after a cut the panel shows a one-click ↩ Undo button that restores the timeline to exactly before the apply (snapshot reconstruction); Cmd+Z is the guaranteed fallback. Tell the user both. A long scan or cut can be halted with the panel's Stop button.

Preconditions

If a tool returns "panel not connected" or "no active sequence": Premiere must be open with a project and a sequence, the OpenCutAgent panel open and Connected, and a claude session running in the project. Relay concisely; don't retry-loop.

Relationship to the rest of OpenCutAgent

This is one of the mcp__premiere__ppro_* tools — see the premiere-edit skill for the full toolset (ppro_get_timeline_state, trim, gaps, transcript-based silence/filler, retake removal) and the read → propose → confirm → execute → verify discipline. For "tighten this / remove dead air" reach here first (loudness, free); for "cut the ums" or "remove duplicate takes" use the transcript/retake tools.