Auto cut audio for Premiere Pro
Every other silence remover on this search is an extension that installs into Premiere. This one does not go into Premiere at all. It runs in a browser, finds the pauses, and hands you an XML that Premiere opens as a finished sequence — your rushes never move and nothing is re-encoded.
In short: upload, check the pauses it found, pick your sequence frame rate, download the Premiere XML, then Fichier > Importer. You relink the media once and the cuts are in place. Free tier: 5 min and 200 MB per file, no account.
Nothing installs. That is the whole product.
The silence removers that rank for this query are panels: you download a ZXP, you run an installer or drop it in the extensions folder, you restart Premiere, and a panel appears next to your timeline. When it works, it is a nicer experience than ours. It is genuinely more integrated. We are not going to pretend otherwise.
It just requires that you be allowed to install it. Which rules out more people than the plugin makers like to admit:
- The agency machine. IT owns the admin password and a ticket to install a third-party Adobe extension is not getting approved this week.
- The client’s Mac. You are freelancing in their suite for three days. You are not installing panels on someone else’s edit machine, and you shouldn’t.
- The version you are pinned to. The project was started on an older Premiere and nobody is updating it mid-delivery. Extensions come with a minimum version. XML import does not.
- The panel that broke. Adobe ships an update, the extension host changes, the panel goes blank, and you wait for a patch. A file you import cannot break this way. It is a text file.
So the trade is honest: they get a panel inside your timeline, we get to work on any machine that has a browser, with any version of Premiere, with nothing to uninstall afterwards.
Why Premiere gets FCP7 XML and DaVinci gets FCPXML
This is the detail that decides whether an import works or does nothing at all, and it confuses everybody, because the two formats have nearly the same name and are not remotely the same thing.
FCP7 XML — the old one, root element <xmeml version="4">, file extension .xml — came out of Final Cut Pro 7 and became the lingua franca every NLE learned to read. Premiere reads it. That is the one Premiere wants.
FCPXML — the new one, root element <fcpxml version="1.9">, extension .fcpxml — is Apple’s modern format for Final Cut Pro X. Different vocabulary, different time model (rationals over the frame rate rather than frame counts), different structure entirely. Resolve and Final Cut read it happily. Premiere does not.
Which is why our export buttons are labelled by editor and not by format. You click “Premiere Pro” and get xmeml; you click “DaVinci Resolve” and get FCPXML. Same analysis, same cut list, two serializers. From src/components/TimelineExport.tsx:
const TARGETS = [
{ id: "premiere", label: "Premiere Pro", format: "premiere" }, // xmeml v4
{ id: "resolve", label: "DaVinci Resolve", format: "fcpxml" },
{ id: "fcp", label: "Final Cut Pro", format: "fcpxml" },
{ id: "other", label: "EDL", format: "edl" },
{ id: "otio", label: "OpenTimelineIO", format: "otio" },
];A plugin vendor cannot tell you this, because a plugin vendor lives inside one editor and has no reason to have solved it. We had to solve it to ship a single button per editor. So here is the rest of what we learned writing the serializer, since it is the stuff that costs you an afternoon:
Declare the source file once, not on every clip
In xmeml, each cut is a <clipitem> and each clipitem points at a <file>. The obvious implementation writes the full file definition into all of them. Do that and Premiere imports the same source once per clip — a 200-cut sequence gives you 200 identical bin items. The fix is to define file-1 in full on the first clipitem and write <file id="file-1"/> on every one after it, which is what Premiere’s own exporter does.
NTSC rates are rationals, never decimals
23.976 is not a frame rate. 24000/1001 is. Store the decimal and your cut points drift by about a frame every 40 seconds — imperceptible at the head of a sequence, obviously wrong by the end of a 20-minute one. In xmeml we write <rate><timebase>24</timebase><ntsc>TRUE</ntsc></rate>, which is how the format expresses it: the integer timebase plus an NTSC flag, never the decimal.
Adjacent cuts get merged back together
Two silences less than one frame apart round to the same frame boundary, and a silence exactly twice the padding long can survive our filter on a float rounding error (0.30 − 0.00 reads as 0.30000000000000071) and then collapse to a zero-length gap. Either way you would get a razor cut in the middle of a sentence that removes nothing. So the serializer coalesces clips whose source ranges are adjacent. It is lossless, and it spares you cleaning up phantom edits.
The round trip, step by step
Nobody documents this, for an obvious reason: if you know how to import an XML, you do not need a plugin. Here is all of it.
1. Upload the file you are actually going to cut
Give us the MP4, not a bounced MP3. We write a video track into the XML only when the source we analysed was a video, so an audio upload produces an audio-only sequence. Detection reads the audio either way — the picture just comes along for the ride, cut on the same frames.
2. Check the pauses before you export anything
The waveform highlights every stretch we are about to remove. Auto-calibration has already picked a threshold from your file’s own noise floor, but you can drag it and re-analyse on the spot. Do this now. Fixing a threshold takes two seconds here and a full re-import later.
3. Set the frame rate to your sequence, then hit “Premiere Pro”
Not your camera’s rate, not the delivery rate — the rate of the sequence you are going to paste this into. We never probe the picture (silencedetect works on audio), so this is the one thing we genuinely cannot infer for you. The download is a
.xml, a few kilobytes, no queue, no render: the timeline endpoint reads the analysis already in the database and templates a string.4. Fichier > Importer, pick the .xml
Premiere creates a bin and a sequence inside it, named after your file with (autocut) appended. Open it. Every cut is already in place, and the timecode starts at 00:00:00:00 non-drop.
5. Relink the media — this always happens
Every clip will be offline and Premiere will ask you to find it. Point at your original file once and the whole sequence resolves; the clips all reference the same source. See below for why we do this on purpose.
6. Clean up like a human
The cuts are straight cuts. Roll the audio edges where a join sounds abrupt, drop a two-frame crossfade on the sharp ones, and delete any edit that removed a pause you actually wanted. This is the part a detector should not be doing for you, and the reason handing you an editable sequence beats handing you a flattened MP4.
The two traps, named
Relink. Your file arrived through a browser upload, and a browser upload does not carry a disk path — we never knew where your media lives. So the XML says file://localhost/your-file.mp4: a bare filename, which always shows offline and always offers a relink. That is a choice, not a bug. The alternative is inventing a plausible path, and a wrong absolute path either fails with no explanation or, on a machine with a similarly named file somewhere, links the wrong media and lets you edit for an hour before you notice.
Timebase. Import an XML written at 25 fps into a 23.976 sequence and Premiere will conform it, but every cut point is now on a fractional frame and the drift accumulates. The head of the sequence looks fine. Minute nine does not. Match the dropdown to the sequence before you export, and if you already imported the wrong one, re-download rather than fight it — the export is free and takes a second.
Delete Pauses, the panels, and us
You have probably already tried Premiere’s own answer before landing here, so let us start there rather than pretend it does not exist.
| Delete Pauses (built in) | Installed extensions | Make Autocut | |
|---|---|---|---|
| How it finds a pause | Transcribes the sequence, cuts between words | Loudness threshold, inside the timeline | Loudness threshold, ffmpeg silencedetect |
| Install needed | None — it ships with Premiere | ZXP / extension, admin rights, restart | None. It is a web page |
| Premiere version required | Recent releases only | A supported minimum, per vendor | Anything that imports FCP7 XML |
| Language dependent | Yes — needs a supported language | No | No. It never listens for words |
| Removes filler words (“um”) | Yes — that is the point of a transcript | No | No. An um is sound, not silence |
| Other editors | Premiere only | Premiere only | Resolve, Final Cut, EDL, OTIO — same analysis |
| J-cuts / L-cuts | No | Some, on paid tiers | No. Straight cuts only |
| Touches your media | No | No | No, on the XML path |
| Price to try | Included | Trial, then a subscription | Free: 5 min · 200 MB · 1 file/day, no account |
The row that matters is the third one from the bottom. Delete Pauses and every panel on this search do one editor. We do the analysis once and serialize it four ways, which is only interesting if you actually move between editors — and if you don’t, a transcript-based tool inside your own timeline is a perfectly good answer and it is already installed.
Adobe documents Delete Pauses under Text-Based Editing: Text-Based Editing — Adobe Help. Extension behaviour above is described by category, not by vendor, because minimum versions and tiers change faster than this page does.
What to set the threshold to
Every tool in this space publishes a dB range and tells you to pick a number in it. That advice is close to useless, because a fixed threshold is the wrong idea. A Zoom H6 in a treated booth and an OBS capture with a laptop fan running do not share a noise floor, and −40 dB means something different on each of them.
So we measure yours first. volumedetect reports the mean volume of your file, we set the threshold to mean − 5 dB clamped into the −50…−25 dB range, and fall back to -35 dB only if the probe returns nothing. That 5 dB gap is the entire safety margin, and it is deliberately timid: on all five files in our public benchmark, auto lands in the lower half of the range and under-cuts rather than risk clipping a word. Which means the tuning advice below is not “set it to X” — it is “start on auto, then nudge”.
| Recording | Threshold | Min. silence | Padding | Why |
|---|---|---|---|---|
| Two-host podcast, treated room, decent mics | Leave it on auto | 0.8 s | 150 ms | Auto-calibration has room to work here: the gap between voice and noise floor is wide, so mean − 5 dB lands in a sane place. This is the case the defaults were built for. |
| Voice-over recorded at home, no treatment | Auto, then +3 to +5 dB | 0.8 s | 150 ms | A prepared read has little dead air to begin with — on our bench, an already-published voice-over gave back 3.3% at defaults. Nudging the threshold up is what finds the rest. |
| Screencast, Loom, OBS, Zoom lesson | Auto, then +5 dB | 1.2 s | 200 ms | Teaching pauses are long and meaningful. Raising the minimum keeps the beat after a question and only kills the “where is that slide” dead air. Extra padding stops the joins sounding clipped over a static screen. |
| Outdoor interview, handheld or lav | Auto, expect to fight it | 1.0 s | 200 ms | Wind, traffic and room tone sit close to the voice. If the tool removes almost nothing, that is the honest answer and no threshold will fix it — denoise first, then come back. |
Defaults are 0.8 s minimum silence and 150 ms padding at each end, so a 1.00 s pause loses 0.70 s and keeps 0.30 s of air. All of this is measured rather than asserted — the main auto cut audio page has 155 threshold measurements on five public-domain recordings you can download and re-run yourself, including the one where the tool honestly fails.
Four things the panels do and we don’t
A comparison table written by the vendor always has the empty boxes on the other side. Ours doesn’t, so here they are.
J-cuts and L-cuts
Offsetting the audio ahead of or behind the picture so a cut lands under the previous shot's sound. It is the single biggest thing between a rough assembly and something that sounds edited, and our serializer lays audio and video in lockstep — every cut we write is a straight cut. You can roll the audio edges yourself once the XML is in, which is exactly what the plugin automates. Doing it in the serializer is on our list; it is offset arithmetic on data we already hold.
Disable instead of delete
Leaving the silences on the timeline as muted or disabled clips so you re-approve each one by hand, instead of rippling them out. Every clipitem we write is enabled TRUE and already tightened. In XML terms this is a small change and we intend to make it.
Track selection on multitrack
Two mics on an interview, and you want the detection driven by one of them. We analyse the mix and write one stereo audio track. For a conversation the mix is usually the right answer — a pause is only a pause when nobody is talking — but you have no way to override it.
Saved presets
We auto-calibrate to each file, which is better than a preset for the threshold and no help at all for the minimum-silence and padding rhythm you prefer. Set them every time. It is a real annoyance and we know it.
If any of those four is the reason you are searching, buy the plugin. It will do the job and this page will still be here when you are on a machine you cannot install it on.
The words, if you are new to this
Half of this vocabulary comes from tape and none of it is explained on the pages selling you the tools.
- Ripple delete
- Remove a chunk and close the gap, everything after it shifting earlier. It is what makes the sequence shorter. A plain delete leaves a hole. Our XML is already rippled — the clips sit end to end.
- Straight cut
- Audio and picture cut on the same frame. Everything we write is one of these.
- J-cut
- The audio of the next shot starts before its picture — you hear them before you see them. Named for the shape the two clips make on the timeline.
- L-cut
- The reverse: the picture cuts away while the previous audio keeps running under it. Standard for interviews, where the answer continues over a cutaway.
- Padding
- The air we hand back at each end of a removed gap. Cut a pause flush and the join sounds clipped, because speech does not start at full volume. Default is 150 ms per side.
- Noise floor
- The level of the quiet in your recording — hiss, air conditioning, the room. If it sits within ~10 dB of your voice, no threshold separates them and no loudness-based tool can help you.
- Room tone
- A recorded chunk of that quiet, kept on purpose to paste under edits. Removing every silence removes your room tone, which is why a heavily cut track can sound like it is stuttering between spaces.
- Timebase
- The frame rate the sequence counts in. 24 with an NTSC flag means 23.976, i.e. 24000/1001. It is not 23.976 and storing it that way costs you a frame every 40 seconds or so.
Cut a file and take the XML
First one is free, no account. You see the waveform, the pauses and the threshold it picked before anything is exported — and the timeline download appears as soon as the analysis is done. You never have to run the render.
Editing in DaVinci instead? Auto cut audio for DaVinci Resolve covers the FCPXML route, the media pool checkbox and the frame-rate lock.
Questions people actually ask
Does the XML route re-encode my footage or lose quality?
▾
No, and this is structural rather than a promise. The XML path never touches your media — it is string templating over a list of in and out points that already exists in our database. No ffmpeg runs, no worker picks the job up, nothing is decoded. Premiere opens your original file off your own disk and plays it. Compare that with our MP4 export path, which really does re-encode to H.264 CRF 20 with AAC at 192 kbps. If quality loss is the thing you care about, take the XML and never touch the render button.
Why does Premiere show every clip offline when I import the XML?
▾
Because we deliberately give it a filename and not a path. We never had your disk path — your file reached us through a browser upload, which strips it. So the XML points at file://localhost/your-file.mp4 and Premiere reports the clip offline and offers to relink. Point it at the original file once and the whole sequence resolves. A guessed absolute path would be worse: it fails silently or, on a shared machine, links the wrong file.
Do I need a particular version of Premiere?
▾
Any version that can import Final Cut Pro XML, which means essentially every version Premiere has shipped for over a decade. The reason we can say that is that we are not an extension. Nothing installs, nothing registers with the CEP or UXP host, nothing breaks the next time Adobe ships an update. If your machine runs Premiere and a browser, you are compatible.
Which frame rate do I pick in the export dropdown?
▾
The frame rate of your Premiere sequence, not the frame rate of the camera and not what you plan to export. Silence detection reads the audio stream, so we never see a picture frame rate and we cannot guess it — you tell us, and we quantise every cut point to whole frames at that rate. Get it wrong and the cuts drift progressively out of sync. We offer 8 rates and the NTSC ones are exact rationals (23.976 is stored as 24000/1001), because rounding that to a decimal costs you a frame roughly every 40 seconds.
How is this different from Premiere's own Delete Pauses?
▾
Delete Pauses works from a transcript. Premiere transcribes the sequence, finds the gaps between words, and removes them — so it depends on the spoken language being supported and on the transcription being right. We measure loudness with ffmpeg silencedetect and never listen for words, which means every language behaves identically and a 90-minute file costs no transcription time. The flip side is real: a transcript knows an “um” is a word and we never will. If your problem is filler words rather than dead air, Delete Pauses or Descript is the better tool and we would rather say so.
Does it handle multitrack sequences — two mics on an interview?
▾
Not properly, and this is our biggest gap. We analyse the mixed-down audio and we write a single audio track into the XML with one stereo clip per kept segment. If you recorded two people on two separate mics, the detector treats a moment where either of them speaks as “not silence”, which is usually what you want for a conversation but gives you no way to say “cut on the host's mic only”. Track selection is on our list. It is not there today.
Can it do J-cuts and L-cuts?
▾
No. Every cut we write is a straight cut: audio and video start and end on the same frame. A J-cut or L-cut needs the audio to lead or trail the picture, which means offsetting the two tracks against each other on purpose, and our serializer lays them out in lockstep. Plugins that sell this as a Pro feature are not lying about its value — a hard cut on a conversation is audibly a cut. Once the XML is in Premiere you can drag the audio edge of any clip past the video edge yourself with the rolling edit tool, which is exactly what the plugin automates.
Can I mute the silences instead of deleting them?
▾
Not today. Our XML writes every clipitem with enabled TRUE and ripples the gaps out, so what lands in Premiere is already tightened. The non-destructive part of our approach is at a different level: your rushes are untouched and the XML is a text file you can delete if you hate the result. But if you specifically want the silences left in place and disabled so you can re-approve each one, we do not do that and a plugin that works inside the timeline does.
What are the limits before I have to pay?
▾
5 minutes per file, 200 MB, one file a day, no account. Pro raises it to 60 minutes and 2 GB with no daily cap. The limit is on what you upload, not on what you get back — no watermark on any tier, no degraded export, no locked features in the XML. Be warned that 5 minutes is tight: a real interview will not fit, so if that is your job you are looking at Pro.
The sequence imported but there is no video track. Why?
▾
You uploaded audio. We write a video track only when the source we analysed was a video file — drop an MP3 and you get an audio-only sequence, which is correct but surprising if you were expecting to relink to your MP4. Upload the MP4 itself and you get both tracks, cut on the same frames.
Nothing to install. Try it on a real file.
Five minutes, no account, no card, no watermark — and the XML costs nothing extra.
Cut my file and get the XML