Editor is our Premiere Pro plugin for the part of a talking-head edit nobody enjoys: getting the A-roll clean. It transcribes your sequence on your own machine, marks the silences, breaths, filler words, profanity, and repeated takes in one pass, shows you every change in the script, and then builds a clean new sequence next to your original. On the recording in the demo below, a 37 minute take came down to 15:44 before any creative editing started.
Everything about it lives on the Editor product page, and it installs through Filmit Studio, our free desktop app, with a 7-day trial. Here is what it does, what it gets wrong, and where it sits next to the other tools in the suite.
What does Editor actually do?
Three stages, and you stay in charge of the last one.
Transcribe on your machine
Point it at a sequence and pick which audio channels hold the voice, so music and sound effects stay out of the transcript. It runs locally with the same Whisper engine our other tools share, using your GPU when there is one. Nothing is uploaded. A recovery pass then re-listens to any stretch that has sound but no transcribed words, so phrases Whisper dropped come back.
Mark everything in one pass
Tick what to remove: silences, breaths, filler words, profanity, and repeated takes. Each kind of mark gets its own colour in the script, and you can change the colours in Settings.
Review, then build
Read the marked script before anything touches your timeline. Click a mark to keep it, select words to cut them yourself, bring words back inside a cut. When you are happy, Build creates a new sequence next to the original, filed under Editor/Cuts.
The review stage is the point. Editor never edits your original sequence, and nothing is built until you have read what it intends to do.
How does it find bad takes?
By listening for the thing every presenter does: starting a sentence, stumbling, and starting it again. When the same run of words restarts within a short window, Editor treats the earlier attempts as bad takes and keeps the last one by default. The window defaults to 25 seconds and four matching words, and both, along with which attempt stays, are adjustable in Settings.
In the demo, one line was flubbed three times before the fourth attempt landed. Editor caught all three and kept the fourth. It also flags stretches of audio that are not speech at all, like the sound of scrubbing a timeline mid-recording, as not in the script, so those go too.
It is rule-based, so it can be fooled. Later in the same recording, a passage where the presenter read a list aloud looked like a repeated phrase, and Editor marked it as a retake when it was really a continuation. That is exactly why the review exists: one click on the mark kept it, and the cut moved on.
How do you review what it marked?
In a panel built for reading, with five controls doing most of the work.
- Relaxed or Aggressive. Relaxed only cuts words that made it into the transcript, so laughs, reactions, and the odd ‘hm’ stay. Aggressive, the default, also cuts the ums and uhs Whisper skipped, marking them straight from the audio.
- Full script or Review mode. Full script shows everything. Review mode shows only the passages where something changes, with a line of context either side, so checking a long recording means scrolling through edits instead of rereading the whole thing.
- Jump to the moment. Every line carries its timecode. Click it, or right-click a line and choose Jump here, and the playhead lands on that take in your original so you can hear it before you decide.
- Cross out. Turn it on to strike through every word that is going, or leave it off to keep the script easy to read while you listen.
- Keep, cut, restore, reorder. Click a mark to keep it, select any words and right-click to cut them yourself, restore a cut you made, or drag lines into a new order before you build.
What about filler words and profanity?
Both run from word lists you control, per language, in Settings.
Do you need AI or an API key?
No. Everything above is rule-based and runs without an AI account, an API key, or any extra cost. The AI is an optional second reader.
AI takes lets Claude or ChatGPT read the script, find abandoned takes the rules missed, and judge the rule's own cuts. It sends text only, never audio. You can use your own key from Settings, which in the demo came to roughly 20 cents a run, or export the script and paste it into a chat you already pay for, then import the plan it gives back.
It is a reader, not an oracle. In the demo the AI correctly rescued a line the rules had cut, and in the same pass it removed a line that was the right way to open the video. It also flagged a stretch it could not make sense of rather than guessing, which turned out to be audio from a clip playing inside the recording. Expect a different read on every run, and treat it as a second opinion you review like everything else.
How is it different from RoughCut and JumpCut?
They answer three different questions, and the demo uses two of them back to back.
- JumpCut cuts silence by audio level. It does not read words at all. In the demo it took the raw 37 minutes down to about 17 before Editor ever opened.
- Editor cleans a linear A-roll by reading it. You already know the story and the order; it removes the takes, fillers, and noise inside it. That pass took the cut from about 17 minutes to 15:44.
- RoughCut builds a narrative. Point it at hours of interviews and it assembles a story from them. Editor will not do that, and it is not trying to.
If you only have one of these problems, you only need one of these tools. Our guide to cutting silence out of long edits covers the JumpCut half step by step.
What else is in the panel?
Four jobs that are not editing but eat editing time, each one click.
What does a real run look like?
The demo is a full, unedited pass on a real recording, and it uses the rest of the suite the way an editor actually would.
Organizer sorted the project bins first. Shortcut closed the stray timeline tabs. JumpCut stripped the silence, 37 minutes to about 17. Editor transcribed that cut, marked it, and after review built a 15:44 sequence. Then the presenter re-transcribed the new sequence and ran a second review, because cleaning a cut is iterative, and the second read catches things the first one could not see.
Before building, the reviewed script can be exported as a document. That turns out to matter more than it sounds: you are rarely editing alone, and handing a client the marked-up script lets them approve what is coming out before you build it.
What are the honest limits?
- Rules misread some patterns. Reading a list or repeating a phrase on purpose can look like a retake. The review catches it, but you do have to review.
- AI takes vary. A different read each run, sometimes better than the rules and sometimes confidently wrong.
- Transcription is not perfect. Whisper occasionally drops a short phrase. The recovery pass catches most of these, but the script can still miss something the audio contains.
- It will not find your story. That is RoughCut's job. Editor assumes the order is already right.
- The Switcher is timed, not smart. Random-length holds make a watchable review cut; they will not cut on the beat or on a reaction for you.
How do you get Editor?
Editor runs in Premiere Pro today, with a DaVinci Resolve version planned. Head to the Editor page, install Filmit Studio (free), and start the 7-day trial. Editor is included with the Filmit Studio subscription alongside every other tool in the suite. Try it on a real recording, and tell us on Discord what it gets wrong; that is what the next version gets built from.
Key takeaways
Silences, breaths, fillers, profanity, and repeated takes are marked together, and nothing is built until you have read the script.
The cuts are rule-based and free to run. AI takes is an optional second reader that sends text, never audio.
Build makes a new sequence under Editor/Cuts. The Switcher keeps a backup copy before it cuts.
JumpCut removes silence by level, Editor cleans a linear A-roll by reading it, RoughCut builds a narrative. Use the one that fits the problem.
