Transcribe an interview and edit by text on a Mac, locally

Editing an interview by text means you transcribe the footage, read the transcript like a document, and cut the video by selecting and removing words. It turns a two-hour sit-down into something you can scan in twenty minutes, and it lets you build a structure on paper before you fine-tune it by ear.

In cinemaOS the whole loop runs on your Mac: transcribe with ⇧⌘T, check the speakers, edit in the Text Editing bin tab, search the transcript for the lines you need, optionally ask J.A.R.E.D. to tighten the cut, then caption. No media, transcript or prompt is uploaded. Below is the workflow, followed by the editorial habits that make interview cuts work.

Why transcribe locally

Interviews are often the most sensitive footage in a project. A subject may say things off the record, a documentary may involve people who agreed to be filmed but not to have their words sent to a third-party service, and a corporate client may have rules about where recordings go.

Local transcription removes that question. cinemaOS uses Parakeet, which covers 25 languages, and Whisper for the rest. Both run on the Mac, so the job works without a network connection and the transcript stays in your project. The trade-off is speed on smaller machines: on an 8 GB Mac, heavy AI jobs run one at a time, so start transcription before you break for coffee rather than while you are also upscaling footage.

If privacy is a central concern on your project, on-device AI video editing and privacy goes into what local processing does and does not guarantee.

Step 1: Transcribe and check the speakers

Select the interview clips and press ⇧⌘T, Transcribe with J.A.R.E.D. You can also do it from the Finisher in Localize ▸ Captions. Leave the language on Auto Detect unless you know the recording mixes languages in a way that confuses detection, in which case set it.

cinemaOS detects up to four speakers. That covers most interview setups: the subject, the interviewer, and one or two others. Once the transcript appears:

1. Name the speakers. Replace generic labels with real names. It makes searching and reading much faster later.
2. Spot-check speaker changes. Short interjections like "right" or "mm" are where speaker detection most often guesses wrong. Fix the ones that matter.
3. Correct names and terms. Proper nouns, technical words and place names are the usual misses. Fix them now, because the same transcript feeds your captions later.

You do not need to proofread every word at this stage. Fix what will affect search and captions; the rest can wait until you know which lines survive the cut.

Step 2: Edit in the Text Editing tab

Open the Text Editing tab in the Sequencer bin (next to Bin and Forge). The transcript is now an editing surface. Selecting text selects the matching media, and removing words removes the matching section from the cut.

A practical order of work:

1. Do a selects pass. Read through once and keep only the answers that could make the film. Be generous. You are removing obvious dead ends, not making final decisions.
2. Build a paper structure. Arrange the surviving answers into the order of your story. This is where text editing is strongest: moving paragraphs is faster than moving clips.
3. Then switch to your ears. Play the assembly through. Text hides tone. A line that reads well may land flat, and a stumble that looks messy may be the most honest moment in the interview.

Treat the transcript as a map, not the territory. The final judgement on every cut is made on the timeline, by listening.

Step 3: Search the transcript for what you need

Long interviews are hard to remember. Use Search Project… to find moments by Word or by Meaning. Word search finds exact phrases, which is ideal when you remember what someone said. Meaning search is for when you remember the idea but not the wording, such as "the part where she talks about leaving home".

Two habits that help:

• Search for the question, not just the answer. If your interviewer asked the same thing twice in different ways, both answers are candidates.
• Search across every interview. If several subjects touch the same theme, find them all before choosing. It is the fastest way to build a section that cuts between voices.

Step 4: Tighten with J.A.R.E.D., on your terms

Once you have a rough assembly, J.A.R.E.D. can do the mechanical tightening. Press ⌘J, set the session mode to Always ask, and type something like:

Tighten this interview to 2 minutes and keep the pauses

Or, if you want a lighter touch:

Remove filler words as a candidate

J.A.R.E.D. runs on the Mac and the edits land on the real timeline as you watch. With Always ask, it proposes before it acts. When it finishes, you get a receipt of what changed, and ⌘Z reverses the whole step. Destructive actions always ask, whatever the mode.

Be realistic about what this is for. It is good at length targets and at removing ums and false starts. It does not know which answer is the emotional centre of your film. Use it after you have made the structural decisions, not before. J.A.R.E.D. also works best with requests written in English today; other languages are understood less reliably.

And keep the pauses. An interview with every breath removed sounds airless and rushed. A half-second pause before an important answer is often the reason it lands. If you want to know more about why an AI's "done" is not the same as a good cut, when the AI editor says done but the cut is wrong covers how to review an automated edit.

Step 5: Caption from the same transcript

When the cut is locked, you already have most of your captions. You can chain the two steps by asking J.A.R.E.D. "Tighten interview, then caption", or add captions yourself: ⌃⌥C adds a caption at the playhead, and Finisher ▸ Localize ▸ Captions holds the full subtitle tools.

Before export, run QC with Strict or Compatible Repair, and check line length and reading speed on the busiest passages. Then export in the format your platform needs: SRT, WebVTT, ASS, SCC, EBU-STL, iTT, IMSC1, EBU-TT-D or FCPXML, or use Burn In for social versions. If you need the interview in another language, the same transcript can be translated locally; translating subtitles on your Mac explains the quality tiers honestly.

Interview editing habits worth keeping

Whatever tool you use, these make interview cuts better:

• Cut on thought, not on sentence. People finish thoughts mid-sentence and start new ones mid-breath. Find the real start of the idea.
• Watch the eyes before a cut. A subject looking down to think and back up to speak gives you a natural edit point.
• Cover jump cuts honestly. B-roll, a cutaway to the listener or a deliberate jump cut all work. A hidden jump cut that changes meaning does not.
• Keep the subject's voice. Tightening is not rewriting. If a cut makes someone say something they did not mean, put it back.
• Read the final transcript once. Before captions, read the cut as text. Repeated words and sentences that do not connect are easier to spot on the page.

FAQ

How many speakers can cinemaOS detect?

Up to four per transcription. For a panel with more people, name the four main voices and correct the rest by hand.

Does transcription work offline?

Yes. Transcription runs locally with Parakeet or Whisper, so it does not need a network connection, and nothing is uploaded.

Will J.A.R.E.D. change my cut without asking?

Not in Always ask mode, which proposes each change first. In any mode, destructive actions ask for confirmation, and every edit undoes with ⌘Z.

Try it on your next interview

Take one interview, transcribe it, and do a selects pass in the Text Editing tab. That alone usually saves an afternoon. cinemaOS runs on Apple silicon Macs with a 60-day free trial, no card required.