Remove ums and uhs. In your browser.

Remove filler words in three steps

1. Choose an audio or video recording, or try the sample. Press Find filler words. The first time, the speech model downloads; the page shows its size and progress first.
2. Read the transcript. Every um, uh and hmm is highlighted and selected for removal. Click a highlight to keep it, click any word to hear that spot, and switch on “like”, “you know” or long pauses if you want those gone too.
3. Press Remove selected, compare the edited preview with the original, and download the audio or video. The transcript is available as a .txt file and as .srt subtitles timed to the edited file.

How it finds filler words

The recording is converted to 16 kHz mono and split into pieces under 30 seconds at its quietest moments, because Whisper reads 30 seconds at a time. Each piece is transcribed with word-level timestamps by Whisper base.en, an English speech-recognition model from OpenAI, running in your browser through transformers.js and ONNX Runtime Web. Words spelled like hesitations (um, umm, uh, er, erm, hmm, mm) are marked as fillers. “Like” and “you know” are only offered when Whisper sets them off with commas, and they start unselected, because “I like it” and “you know the answer” are real speech.

Why some ums are missing from a transcript

Speech-recognition models are usually trained on cleaned-up captions, and Whisper tends to drop hesitations to make the text read well. To keep them, AudioClover starts Whisper with a short prompt that itself contains ums and uhs, which nudges the model to write them down. It still misses some. When “Sounds between words” is on, the tool also marks stretches of voice that sit between two transcribed words with no word on them: often a filler, sometimes a breath or a laugh. Those start unselected, so listen first.

Clean cuts, not choppy ones

Whisper’s word times can be off by a fifth of a second, so each cut edge is moved to the quietest 10 ms moment nearby, never past the middle of the neighbouring word. The two sides are joined with a 10 ms crossfade, the same joining code the silence remover uses, so there is no click. Long pauses can be shortened at the same time, leaving 0.4 seconds of breathing room. Video keeps picture and sound in sync: the same cuts are applied to the frames.

Speed, and what to expect on your device

On an Apple M2 laptop in our tests, an hour of speech took about 7 minutes to transcribe with WebGPU (the encoder runs on the graphics chip) and about 13 minutes with WebAssembly alone, the path browsers without WebGPU use. Phones and older computers are slower. The page shows an estimate before you start and the time left while it works, and the tab stays responsive because the model runs in a background worker. A phone may run out of memory on an hour-long episode; a computer handles it better. The model and the 28 MB engine are downloaded once and then stay in your browser’s storage.