Bleep out swear words in audio and video.

At a glance

How it finds the words

The recording is transcribed in your browser by Whisper, the open speech model, in a version that also gives the time of every word. Whisper writes profanity out in full rather than starring it, so each word of the transcript is checked against a list: the strong words and everything built on them, the common slurs, and, if you tick the box, mild words like damn and hell. The list matches whole words where it has to. "Hell" does not catch "hello", "ass" does not catch "class" or "assume", and "Scunthorpe" is left alone. Words with an everyday meaning, such as bloody or screw, are not on the list at all, because a DIY video or a sermon would come out bleeped throughout; add them under Extra words if you need them.

You are not limited to swearing. Every word in the transcript is a button. Click a name, a street, a company or a phone number and it is bleeped the same way. Click a marked word to let it through, for example a quoted title. Shift-click any word to hear it in context before you decide.

Why the bleep lands on the word

Word timestamps are a guide, not the cut. Whisper's start times are close, but its end times run early, so a bleep placed exactly on them would let the last syllable through: "fu-" bleeped and "-cking" left audible, which is worse than no bleep at all. The page measures the recording's own background level and extends each bleep to where the word's sound actually stops, without running into the next word. It errs on the wide side, because clipping the edge of the next word is a smaller mistake than leaving part of the swear audible. In our test recordings, with swear words laid into running speech at known times, every one was found, nothing else was marked, and the loudest part of a word left outside its bleep was 28 dB down, which you cannot hear under the tone. The cost is that a bleep can take in up to about a quarter of a second of the following word. Shift-click to check any you are unsure of.

The bleep is a 1 kHz tone, the broadcast standard, set to the level of the word it replaces rather than a fixed loudness, so it does not jump out at the listener. Silence is the other option, the usual choice for YouTube and podcasts that do not want the tone. Both fade in and out over 5 ms, so there is no click at either end.

Video stays in sync

Nothing is cut out; each word is replaced in place, so the clean audio is exactly as long as the original. For a video, the picture is copied as it is, with no re-encoding and no quality loss, and the new sound is written beside it as AAC. The lips move on the same frames they always did. That makes this the quick route for a clip with one or two slips going to YouTube, TikTok, Instagram or a school or work channel.

Where it falls short

• Mumbled and shouted words. If Whisper does not hear a word clearly it may write something else, and a word it does not write cannot be matched. Listen to the result before you publish, and click any word it misheard.
• Songs. Singing over loud music is harder to transcribe than speech. Separating the vocal first with the stem splitter helps, and you can then bleep the vocal stem and mix it back.
• Other languages. Whisper transcribes many languages, but the built-in list is English. For another language, type the words under Extra words exactly as Whisper writes them.
• Overlapping speakers. When two people talk at once, the bleep covers both of them for the length of the word.

Related cleanup

To take out the ums and uhs rather than the swearing, use remove filler words, which cuts them and closes the gap. For a transcript or subtitles of the clean version, run it through audio to text or add subtitles to video. To even out the level before you publish, use the loudness normalizer.