Remove Filler Words from Video

Upload a video — on-device Whisper AI reads every word, finds your ums, uhs, and likes, then lets you download a clean cut. Nothing ever leaves your browser.

1. Drop your video

🎬
Click or drag & drop a video here
MP4 · MOV · WebM — processed 100% locally, never uploaded

How it works

This tool combines two powerful browser technologies — Whisper AI for speech recognition and ffmpeg.wasm for video editing — entirely on your device.

Step 1 — AI transcription Whisper (OpenAI's open-source speech model) runs in your browser via transformers.js and WebAssembly. It transcribes your video and produces word-level timestamps.
Step 2 — Filler detection Each word token is matched against your chosen filler list (um, uh, like, you know, so, er, ah, literally, basically, actually, right). Near-duplicates and soft matches are caught too.
Step 3 — Segment extraction ffmpeg.wasm cuts the video into keep-segments, skipping every filler interval. A small configurable pad (default 50 ms) prevents jarring cuts.
Step 4 — Concatenation All kept segments are joined with ffmpeg's concat demuxer and exported as a WebM file, ready to download — no re-encoding of the audio required where possible.

Privacy: Your video never leaves your computer. The Whisper model downloads once from Hugging Face CDN and is cached in your browser for future use.

Frequently asked questions

How accurate is the filler word detection?
This tool uses OpenAI's Whisper small model (244 M parameters), which achieves near-human accuracy on clear English speech. Filler words like "um" and "uh" are reliably detected when the audio is reasonably clear. Background noise, heavy accents, or very fast speech can reduce accuracy — you can always uncheck false detections before exporting.
What video formats does this support?
Input: MP4, MOV, WebM, and most formats your browser can decode (MKV, AVI, M4V often work too). Output is always WebM (VP8 video + Opus audio), which plays in all modern browsers and can be converted to MP4 with any video tool if needed. For the best output quality, use MP4 H.264 as your input.
Why does the first run take so long to start?
The Whisper small model is about 150 MB and downloads from Hugging Face on first use. Once cached by your browser, subsequent runs start immediately. Analysis speed depends on your CPU — typically 30–90 seconds of real-time per minute of video. WebGPU-enabled browsers (Chrome 113+, Edge 113+) can be significantly faster.
Can I use this on long videos?
Yes, but keep RAM in mind. The entire video is loaded into browser memory for processing. Most modern computers handle videos up to ~500 MB comfortably. For longer recordings, consider trimming them into shorter segments first, processing each one, then joining the results.
What does the "padding" setting do?
Padding adds a small buffer (default 50 ms) before and after each removed filler word. Without it, cuts right at the word boundary can sound robotic because the preceding breath or trailing consonant gets clipped. A 50–100 ms pad produces more natural-sounding edits. Set it to 0 for maximum time savings, or increase it for smoother audio flow.
Is my video data private?
Completely. Your video file is processed entirely in your browser using WebAssembly — no bytes are sent to any server. The only external request is the one-time download of the Whisper model weights from the Hugging Face CDN. After that, the tool runs 100% offline.