Podcast editing, inside Premiere.

Each speaker’s mic is read on its own, twenty milliseconds at a time. The planner works out who holds the room in every slice, then writes the switch list to your sequence. Run it twice on the same footage and you get the same cut, to the frame.

1. Extract
2. /Detect
3. /Normalise
4. /Plan
5. /Write

1. 01

   Extract

   ffmpeg pulls each speaker’s mic down to mono 16 kHz, the only sample rate the detector needs and the fastest thing to read.

2. 02

   Detect

   Frame-by-frame RMS at 20 ms. The gate is derived from the track’s own 95th percentile, so a quiet room and a loud one get the same treatment. A 200 ms hangover bridges the gap between words instead of chopping mid-sentence.

3. 03

   Normalise

   Each track is divided by its own peak. Now the loudest speaker is the one talking, not the one with the hottest gain.

4. 04

   Plan

   At every boundary the planner asks who should be on screen, then makes them prove it. The candidate has to lead for 70% of the voiced time in the next 0.75 s before the shot is allowed to change.

5. 05

   Write

   The sequence is razored at each cut point and the pieces that aren’t the chosen camera are removed, so the camera underneath shows through. It runs inside a named Premiere undo group.

Analysis frame

20 msnon-overlapping

Minimum shot

1.5 s250 ms debounce

Dominance vote

70%over the next 0.75 s

Writes backRazor cuts on every video track at each switch point, then the losing pieces removed. The camera you want is the one left visible.

Dead air is where nobody at all is talking. One person pausing does not count. Every selected track is gated on its own, and only the spans quiet across all of them can be removed. You see them shaded red on the waveform before anything is cut.

1. Decode
2. /Gate
3. /Intersect
4. /Pad
5. /Ripple

1. 01

   Decode once

   Each track is read to mono 16 kHz and reduced to a frame-RMS envelope plus an 800-bucket waveform. Everything after this is arithmetic.

2. 02

   Gate per track

   Frames under the noise threshold are silent. Runs shorter than the minimum are ignored, because a breath is not a gap.

3. 03

   Intersect

   Only spans silent on every selected track survive. This is the whole idea: one mic going quiet means nothing, all of them going quiet means dead air.

4. 04

   Pad

   Each region is pulled in by 120 ms at both ends, so the cut lands in the silence rather than on the first consonant.

5. 05

   Ripple

   Regions are removed back to front, so earlier edits never move the timecodes of later ones, and every video and audio track ripples together.

Noise threshold

0.020 RMSadjustable

Minimum silence

500 ms120 ms padding

Price

Freenever counted

Writes backA ripple delete across every video and audio track, applied end to start, with a length check afterwards.

Whisper runs locally against each mic you select and returns the transcript word by word. Every word is checked against your list, and you get the hits in a review table before a single frame changes. Untick the false positives, pick what happens, apply.

1. Transcribe
2. /Match
3. /Pad
4. /Review
5. /Apply

1. 01

   Transcribe

   whisper.cpp with the small English model, forced to one token per segment so every word carries its own start and end time. The model downloads once, then never again.

2. 02

   Match

   Each word is normalised to letters and digits and checked against your list. Fifteen stems ship by default; add your own and they persist between runs.

3. 03

   Pad

   80 ms is added either side of every hit, because a word starts a fraction before the transcript says it does.

4. 04

   Review

   Hits land in a table with their timecode and which mic they came from. Nothing touches the timeline until you have been through it.

5. 05

   Apply

   Mute silences that piece of the mic. Bleep does the same and lays a 1 kHz tone over the top on its own Bleep track. Cut ripples the region out of the whole sequence.

Model

whisper small.endownloads once

Hit padding

80 mseach side

Wordlist

15 stemseditable

Writes backMute and bleep leave the sequence exactly as long as it was. Cut ripples every track together. Either way you review first.