Remove filler words: every um, uh and long pause.

At a glance

How it finds an "um"

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 is trained to write tidy text and often leaves fillers out of a plain transcript, but with word timing on it writes most of them: in our test set, eight of nine ums and uhs came back as words, each starting within about a tenth of a second of where it really did. Every word on the filler list (um, uh, ah, er, erm, hmm and their stretched spellings) is marked and ticked.

Whisper sometimes hears a long "mmm" as a real word ("bum" was one in the tests). So a second pass looks at the short words that stand apart, between commas or pauses, and measures their pitch: a filler is one held note, while a real word moves. Those are marked as possible fillers and left unticked, because "so," and "well," at the start of a sentence can sound much the same, and cutting a real word is a worse mistake than leaving an um in. One case in nine was missed entirely: an "uh" Whisper folded into the word before it. Listen through the result before you publish it.

Why the cuts sound clean

Word times are only used to find a filler. The cut itself is placed on the audio: the page measures the recording's own background level, starts the cut where the filler's sound begins and ends it where the sound has died away, which matters because Whisper's word ends run early and an "um" trails off in a quiet hum. Each edge moves to the quietest sample within 5 ms, and the two sides are joined with a 10 ms equal-power crossfade, so there is no click. The silence either side of the filler is trimmed so that the pause left behind is at most a quarter of a second: taking out an "um" should not leave a hole where it was.

Long pauses are handled separately. A gap of more than a second between words is shortened to half a second, but only where the gap is really quiet, so a word Whisper missed is never cut out with it. Untick "Shorten pauses" to keep your timing exactly.

Podcasts, lectures and videos

For a podcast or an interview, export each speaker's track or the final mix and run it here before mastering. WAV and FLAC come out at the file's own sample rate and bit depth, so nothing is lost before you level it with the loudness normalizer. For a video, the picture is cut at the same places as the sound so lips stay in sync. That means the video is re-encoded (H.264 at high quality), unlike the site's other video tools, which copy the picture untouched. A jump cut where a filler was is normal in talking-head videos; in a screen recording it is usually invisible.

What it can't do

It removes sounds, not habits: "like", "you know" and "so" are real words to the model, and the page will not cut them, because whether one is filler depends on the sentence. Use the transcript to find them and the audio editor to cut them by hand. When two people talk over each other, one person's filler may sit under the other's word, and those cuts are best left unticked. Transcription runs on your computer, so a long file takes a while: roughly the length of the recording on a laptop without a fast GPU. Other languages work, but the filler list is English-led; Spanish "eh" is found, other languages' fillers may not be.