Filler Word Remover

Remove filler words from audio and video, without uploading it

Drop in a podcast episode, interview, lecture, or talking-head video, and the ums and uhs are cut in one go. You get a list of every filler with a few words of context, so you can keep any that belong, hear each one, and trim long pauses too. Then export the cleaned-up video or audio. Everything runs in your browser: the file is never uploaded, there is no account, and there is no watermark.

Cloud filler removers make you upload the whole recording before they can find a single um, and for a long episode the upload is often the slowest part of the job. This tool skips it.

How to remove filler words

1. Pick the spoken language, then drop a video or audio file onto the tool, or record one with your microphone or camera.
2. The speech model transcribes it on your device and marks every filler and hesitation. The first run downloads the models once; your browser keeps them, so the next file starts right away.
3. Every filler found is already cut. Untick the ones you want to keep, and press Play on a row to hear two seconds before it through to just after.
4. Tick Shorten pauses over 0.8 s if you want the dead air trimmed as well.
5. Export the video as an MP4, or the sound as M4A, MP3, or WAV.

How it finds fillers on your device

Fillers in the transcript. Whisper, an open source speech model, writes down what was said with a time for every word. Speech models are trained to tidy fillers away, so for English the tool nudges Whisper to keep them in the text, then marks um, uh, uhm, erm, er, ah, hmm, and mm.

Hesitations the transcript missed. Plenty of fillers still go unwritten. A voice activity detector (Silero VAD) checks the audio for short stretches of speech sound that no transcribed word covers, and each one becomes a hesitation row you can keep or cut like any other.

Cuts between words, not inside them. For English, when your browser can use the graphics card, a second model (wav2vec2) lines every word up with the audio and catches fillers sitting in the gaps, so the cuts land in the quiet between words.

Private by design, and free because of it

There is no server doing the work, so there is nothing that could see your recording. That matters for unreleased episodes, client calls, internal meetings, and anything under an NDA. The models download once and stay cached in your browser: Whisper Base is about 200 MB, or about 77 MB on browsers that run it without the graphics card, the voice detector is about 1.2 MB, and the English word aligner is about 67 to 90 MB depending on your graphics card. They are the same models our text-based video editor uses, so if you have run it in this browser, they are already there. Because nothing runs on a server, there is no bill per minute to pass on: the tool is free, with no limit.

Honest limits

• It misses some fillers and flags the odd short word, so listen before you export. On a computer without a usable graphics card, and on phones and tablets, it finds fewer.
• “Like”, “so”, and “you know” are never cut, because they are real words too often.
• English works best. Other languages are transcribed fine, but fewer fillers are found.
• iPhone is not supported yet. iPad and Android work for recordings of a few minutes.
• Your cuts live in this tab only. Export before you close it; the page asks before you leave with cuts you have not exported.

Need more than fillers?

To cut whole sentences, false starts, or the words this tool leaves alone, open the recording in the text-based video editor: delete words from the transcript and the video follows. For just the words, the audio transcription tool gives you a full transcript and subtitles.

Frequently asked questions

How do I remove filler words from a video?

Drop the video onto the tool and pick the spoken language. The speech model transcribes it on your device and every um, uh, and hesitation it finds is cut straight away. Check the list, untick any you want to keep, press Play on a row to hear the moment, then export the video as an MP4 or the sound as M4A, MP3, or WAV.

Which filler words does it remove?

Um, uh, uhm, erm, er, ah, hmm, and mm, plus hesitations: short stretches of speech sound that no transcribed word covers, which is often a filler the transcript left out. It never removes "like", "so", or "you know", because those are real words far too often to cut in one click. To cut them, open the recording in the text-based video editor and delete them by hand.

Does it catch every um?

No, and no tool should promise that. In our English tests on a computer that can use its graphics card, it found about 2 in 3 to 7 in 10 fillers, and about 4 in 5 of the things it flagged really were fillers. Without the graphics card it found about 1 in 3. That is why every filler gets its own row with a checkbox and a Play button: listen to the ones you are unsure about before you export.

Is my audio or video uploaded?

No. Your file never leaves your device. Transcription, filler detection, cutting, and export all run in your browser. After the one-time model download, nothing is sent anywhere, which you can check in your browser's network tab.

Is it really free?

Yes: free, unlimited, no signup, and no watermark. The AI runs on your own computer, so there is no server bill per minute of audio to pass on to you.

Can it remove long pauses too?

Yes. Tick Shorten pauses over 0.8 s and every silence of 0.8 seconds or more between words is trimmed, keeping about 0.15 seconds of silence on each side so the speech still breathes. Untick it to put the pauses back.

Which files can I use, and what do I get back?

Video (MP4, MOV, WebM, MKV) and audio (MP3, M4A, WAV, and more) up to 4 GB, or record straight from your microphone or camera. A video exports as an MP4 at its original resolution, or as audio only; an audio file exports as M4A, MP3, or WAV. Every cut gets a tiny fade so the seams do not click. In Chrome and Edge on a computer the export streams straight to your disk; other browsers build it in memory, so exports estimated over about 750 MB on a computer, or about 300 MB on a phone or tablet, are refused before they start.

Does it work in other languages?

Pick the spoken language before you drop the file. Filler detection is tuned for English: in other languages the transcript is still searched for fillers and hesitations are still marked, but fewer are found, and we have not measured how many yet.

Does it work on iPhone?

Not yet. On iPhone this page crashed within seconds of starting, so it is switched off there until that is fixed. It runs on iPad and Android phones, more slowly and with fewer fillers found, because the speech model runs on the processor. A computer with Chrome or Edge gives the best result.

How long does it take?

On a computer that can use its graphics card, a 38 second clip took about a minute the first time, most of it the one-time model download, and about 7 seconds once the model was cached. Exporting that clip took about 4 seconds as an MP4 and about 2 as audio. Speed depends on your device.

Are the models open source?

Yes. Whisper Base (OpenAI) is released under the Apache 2.0 licence, the Silero voice activity detector under MIT, and wav2vec2 Base 960h (Meta), the word aligner used for English on the graphics card, under Apache 2.0. Whisper and wav2vec2 download once from their public model pages and the voice detector from this site, and all three stay cached in your browser.