Editing the words before you touch a picture

The method that changed interview editing

For anything built on speech — an interview, a podcast with cameras, a vlog, a lecture, a tutorial — the fastest route to a structure is to stop looking at pictures. Transcribe everything, edit the transcript like a document, and only then go to the timeline.

This is not a shortcut invented by software. Documentary editors have done it with paper transcripts and scissors since the 1960s. What changed is that transcription is now free and instant, and several editors can cut the timeline directly from the text.

Why text is faster than video for this

Three reasons, and they compound.

• Reading is roughly four times faster than listening. A 40-minute interview is a 15-minute read. You can hold the whole thing in your head at once, which you cannot do at playback speed.
• Text is searchable. Finding the sentence about cost takes two seconds, not ten minutes of scrubbing.
• Rearranging text is free. Moving a paragraph costs a keystroke. Moving a section of timeline costs a minute and risks breaking the audio underneath.

The cost is that you are working blind to performance. The best-worded answer is sometimes not the best-delivered one. So the transcript decides the structure, and the timeline decides the takes.

Getting a transcript for nothing

Whisper, OpenAI's open-weights speech model, is the workhorse. whisper.cpp runs it on an ordinary laptop CPU with no GPU and nothing leaving the machine, which matters when the audio is confidential.

# convert to 16 kHz mono, then transcribe with word timings
ffmpeg -i interview.mov -ar 16000 -ac 1 interview.wav
./main -m models/ggml-medium.bin -f interview.wav \
       -l en --output-srt --max-len 42

The medium model is a reasonable accuracy-for-time trade on a laptop; small is about three times faster and noticeably worse on accented speech.

Other free routes: YouTube's automatic captions, by uploading unlisted and downloading the track; Resolve's built-in transcription on versions that have it; Kdenlive's speech-to-text, which uses Whisper or Vosk underneath.

Accuracy caveat, and it matters here more than for captions. Recognition degrades on accented English, code-switching mid-sentence, proper nouns and technical vocabulary — which is most of the speech this readership works with. For a caption track that error is a wrong word on screen. For a paper edit it is worse: you may cut out the sentence you needed because the transcript says something else. Always play the audio of any line you are about to cut on the strength of its text.

The paper edit itself

1. Get a transcript with timecodes. Paste it into any document editor.
2. Read it once, end to end, marking nothing.
3. Read it again, deleting everything you are certain you will not use. Do not agonise; you are removing the obvious 60%.
4. What is left, cut into blocks and group by subject, not by the order they were spoken.
5. Order the blocks into an argument. Say the argument out loud in one sentence. If you cannot, the structure is not finished.
6. Print or export the result with its timecodes. That is your assembly plan.

The discipline that makes this work: keep the timecodes attached to every line, all the way through. A paper edit without timecodes is a lovely document you cannot execute.

Text-based editing inside the editor

Several editors now do this without leaving the timeline. Resolve 19 and later has a Text Based Editing panel: transcribe a clip, and deleting a sentence in the text pane deletes it from the source, with the selected text appended to the timeline. Premiere has the same feature. Descript is built entirely around it.

The mechanic is genuinely fast for assembling a first cut of a talking head. Two honest limits. Word-level trim points land on word boundaries in the transcript, which are approximate — you will still be trimming by frame afterwards to remove breath and lip smack. And the picture is being cut wherever the words are cut, which produces a wall of jump cuts that the next block's techniques exist to handle.

The radio edit

The result of all this is what editors call a radio edit: the entire piece cut for sound only, with the pictures ignored or left as whatever happened to be under the words. It should be listenable end to end with the screen off, and it should make its argument.

Get the radio edit right and the picture edit becomes a covering job — choosing what the viewer looks at while they listen — which is a far easier problem than building both at once. Get it wrong and no amount of b-roll rescues it.

What this method is bad at

Anything where the picture carries the meaning. A dance piece, an action sequence, a silent montage, a music video, a process film where somebody's hands are the story. Transcript-first editing is a tool for speech-led work, and applying it to picture-led work produces something that sounds coherent and looks like nothing.

Today

Transcribe any five-minute recording you own and cut the transcript down to two minutes in a text editor, keeping the timecodes. Time yourself. Then estimate how long the same reduction would have taken in a timeline.