WORK IN PROGRESS LipScribe is an early experiment: its lip-reading models are still being trained and get many words wrong. Use it for the gist, and check every word.
Real LipScribe output on this clip, read with the sound off. Video: “The Standard Model”, Fermilab, CC BY 3.0 (cut, cropped, muted).
LipScribe finds the faces in your video, follows the speaker’s mouth and reads their lips with an AI model that runs entirely in your browser — English speech, and Spanish in beta. You get a timed transcript in which every word shows how likely it is to be right. Nothing is uploaded. Free.
Each spoken language has its own lip-reading model (English 169 MB, Spanish 169 MB). It downloads to your browser once, the first time you use that language, and stays cached — after that LipScribe works offline. Everything runs on your own processor; your video never travels.
MP4, MOV or WebM from your phone or computer — one clip or several. It opens in the browser tab; nothing is sent anywhere.
Faces are found and followed through the clip automatically. With several people on screen, tap the one you want; the app shows exactly which mouth it is reading.
Timed lines, with every word shaded by its chance of being right. Tap a word for what else it could be. Export TXT, SRT or VTT subtitles, or JSON with every alternative.
Many sounds look the same on the lips — “pat”, “bat” and “mat” are one mouth shape — and even trained human lip readers miss a lot. So instead of pretending to be perfect, every word gets a probability and you see the other likely readings. Expect the gist, not a verbatim record. Figures measured on our test set: 67 real-world clips, 60 minutes.
Thank you all for coming tonight
Before reading, the app checks the face size, head turn, detection coverage and frame rate, and tells you if a clip is likely to be hard.
“We don't upload your video” is a promise every service makes. LipScribe is built so you don't have to take it on faith:
Network tab. Read a clip — you'll see the model download to you, and nothing leave.A video recorded with the mic off, or shared without sound: get a rough idea of what the speaker said.
Wind, traffic, music or a crowd over the voice. Noise doesn't reach the lips, so they still carry the words.
Draft captions for your own recordings when there is no usable audio track — a starting point to correct, not finished subtitles.
Family films and archive reels whose sound was lost or never recorded, read on your own machine.
Please use it responsibly: on videos you made, or footage you have the right to work with, and with respect for the people in it.
No. Your browser opens the file, the AI models run in your browser's memory, and the transcript is saved to your downloads folder when you export it. There is no server component: we couldn't see your video if we wanted to.
English, and Spanish in beta. Pick the spoken language in the app before reading: it's the language spoken in the video, and it's separate from the interface language (English, Spanish, German, French or Chinese). Spanish is a little less accurate than English for now.
On varied real-world English clips it gets about two words in three right; on a clear, front-facing studio speaker, about 85%. Spanish (beta) gets about 65% right. Results vary a lot with the footage, which is why every word carries a probability and you can see the other likely readings of each word and each line. Expect the gist rather than a verbatim transcript, and expect trouble with names, numbers and rare words.
Each word's percentage is its chance of being right, estimated from what the model saw around that word. White words are likely (80%+), amber ones unsure, and underlined coral ones doubtful (under 50%). Line readings show how the probability is shared among the readings the decoder kept. On our test set, words shown at 80–100% were right about 93% of the time.
One person facing the camera, face reasonably large, good light, 25–30 fps. MP4, MOV and WebM all work in recent browsers. Several people in the shot are fine — you pick whose lips to read.
The first time you read a language, your browser downloads its lip-reading model (English 169 MB, Spanish 169 MB) and keeps it in its cache. After that it starts right away, even offline. On a laptop CPU, reading runs about 1.6× faster than real time — a one-minute clip takes about 40 seconds; phones and older computers take several times longer.
It's free. The open lip-reading models it uses (Auto-AVSR for English, ViSpeR for Spanish) are licensed for research and non-commercial use only, so LipScribe is too: personal, educational and research use are fine; commercial use would need models trained on rights-cleared video. If LipScribe helped you, there’s a pay-what-you-want donation jar.