Rescroll

Remove filler words from audio, podcasts and voice-overs.

Drop in an MP3, M4A or WAV. Every um and uh is marked on the transcript and the waveform; check them, keep any you like, and download the recording back in the same format, tightened. It never leaves your device.

  • 100% free
  • No watermark
  • No sign-up
  • You check every cut
  • Files stay on your device

Loading the tool…

It runs in your browser. Nothing is uploaded.

How it works

How to remove filler words, free, in three steps.

  1. 1

    Drop in your video or recording

    MP4, MOV, WebM, MP3, M4A or WAV, from a phone, a camera, a screen recorder or a podcast mic. It opens in your browser; nothing is uploaded.

  2. 2

    Find the fillers

    The speech model writes down everything that's said, ums and uhs included, and every filler is marked in the transcript. The first time, the model downloads once and stays in your browser.

  3. 3

    Check them, then download

    Play any filler and keep the ones you want. See how many there were and how often, then download the file with the rest cut out: an MP4 for video, the same format for audio.

Why this one

Free, with nothing held back, and nothing uploaded.

Finds the ums a transcript leaves out

Speech models are trained on tidy transcripts, so they quietly drop the ums and uhs. This one is prompted to write them down, which in our tests took it from finding almost none of them to finding most.

You decide every cut

Each filler is marked in the transcript with its time. Play it, keep it or cut it, one at a time or a whole kind at once, and play the result before you download.

Cuts that sound like a breath

Each cut runs from the quiet just before the filler to the quiet just after it, found in the sound itself, with a 10 ms fade so nothing clicks. Keep the pause, or shorten it.

Free, private, full quality

The model runs on your own device, so your recording is never uploaded. Videos come back as an MP4 at their own resolution, audio in the format it came in, with no watermark.

Your file is opened and processed by your own browser. It is never sent to a server, so it stays private, and there’s no upload to wait for.

Removing filler words from a podcast or voice recording

In audio there's no picture to hold attention, so the ums stand out: a podcast host who hesitates before every answer, an interview that wanders, a voice-over read in one take. Editing them out in Audacity or GarageBand means finding each one on the waveform by ear.

Here the recording is transcribed in your browser and every filler is marked on the transcript, with the cuts shaded on the waveform. Play any of them, keep the ones that matter, and turn on repeated words to catch the “I I think”s too. It works on interviews with several people: it finds the fillers whoever says them.

An MP3 comes back as an MP3, an M4A as an M4A and a WAV as a WAV, with only the cut parts gone.

Rescroll

Fillers out. Now the rest of the edit.

Rescroll's AI video editor takes the whole talking-head edit from a sentence. “Cut silences, remove bad takes” takes out the pauses, the ums and the false starts and keeps your best read of each line. Then ask for zooms, captions and graphics in the same chat.

  • Pauses, ums and false starts cut from one sentence
  • Your best take of each line kept
  • Zooms, captions and motion graphics by asking

Questions

The filler word remover: your questions, answered.

How do I remove filler words from a podcast?

Drop the episode's MP3, M4A or WAV on this page and press Find filler words. Check the fillers marked in the transcript and on the waveform, click any you want to keep, then press Cut and download the episode in the same format. Free, and the recording stays on your device.

Will the audio quality drop?

A WAV stays lossless: only the cut parts are taken out. MP3 and M4A are compressed formats, so the kept audio is encoded again, at a high bitrate. If you need it untouched, use a WAV.

Can it tell who said each filler?

No. It finds the fillers in the whole recording, whoever says them, so an interview's guest and host are both cleaned up. Keep any you'd rather leave in by clicking them.

Which filler words does it find?

Always the hesitation sounds: um, uh, er and ah (a hummed hmm and a British erm count as um). If you ask, also like, you know, I mean, so, basically, actually and literally, but only where they're set off from the sentence the way fillers are, so “I like it” and “it looks like rain” are left alone. And if you ask, repeated words and stutters, like “I I think” and “th- the”: the speech model writes down about a third of those, so it catches some, not all.

How accurate is it?

We measured it on 46 minutes of unscripted meeting speech from eight speakers, with every um and uh marked by hand: it found two thirds of them running on a computer's processor and nearly 4 in 5 on its graphics card, and more than 9 in 10 of the words it marked were real fillers. The ones it misses are mostly the shortest, a quick uh tucked between two words. Everything it finds is shown to you before anything is cut.

Why doesn't a silence remover take out the ums?

Because an um isn't silence: it's said at about the same level as the words around it, so anything that cuts by loudness keeps it. This tool listens to the words instead. For the long pauses between sentences, use the free silence remover on this site; the two work well one after the other.

Will the cuts sound choppy?

They're placed in the quiet before and after each filler, found from the sound rather than guessed from the transcript, and the sound fades over 10 milliseconds into and out of every cut, so nothing clicks. The pause around a filler stays as it was, or choose a short pause to tighten the sentence. In a video the picture jumps at each cut, as in any jump-cut edit.

Does it work in languages other than English?

Filler words are found in English only: it's the only language we measured. Recordings in other languages are still transcribed, and repeated words are still found in them, but their own fillers aren't.

Why does it download a model the first time?

Because the listening happens on your device, the speech model has to be on your device too: OpenAI's Whisper, downloaded once from Hugging Face, usually 110 to 140 MB with its engine (280 to 350 MB for High accuracy; the page shows the size for your device first), and kept in your browser for next time. Only the model is downloaded; your file isn't sent anywhere.

How big a file can it handle?

There's no limit of ours: only the sound is kept in memory while it listens, about 3.8 MB a minute, and long recordings are transcribed in pieces. What grows is the time. On a MacBook Pro's graphics card (M4 Max), a 10-minute recording took about a minute to listen to; older computers and phones are slower, so keep to shorter files there.

Is this filler word remover really free?

Yes. There's no price, no trial, no account and no limit on how many files you do. The filler word remover runs on your own device, in your browser, so it costs us nothing to run and we don't charge for it.

Does it add a watermark?

No. The file you download has no watermark, no logo and no intro or outro added. It's your file, changed only in the way you asked.

Is my video or audio uploaded anywhere?

No. The file is opened and processed by your browser, on your device, and never sent to a server: not to us, not to anyone. You can check: watch the Network tab of your browser's developer tools while it works, and you'll see the file is never sent anywhere.

Which browsers does it work in?

Current versions of Chrome, Edge and Safari, and Firefox on a computer: any browser with the WebCodecs API for encoding video and sound. Some files, like iPhone HEVC video, need Chrome, Edge or Safari to decode. A computer is much faster than a phone for long videos.