Add a video. Pick the speaker. Read.
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. You get a timed transcript in which every word shows how likely it is to be right.
Your video, as it is
MP4, MOV or WebM from your phone or computer — one clip or several. It opens in the browser tab; nothing is sent anywhere.
Choose whose lips to read
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.
A transcript you can check
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.
Lips alone, or lips and sound together
By default LipScribe reads only the lips. That is what you want for a muted clip, or when the sound is too poor to use.
For videos with sound, tick Also listen to the sound. A larger model then reads the lips and the voice together: close to 9 words in 10 right, against about 2 in 3 from the lips alone. In English, because it watches the mouth, it keeps to the person on screen even when someone else talks over them. The Spanish model holds up in crowd or street noise, but not when another voice is as loud as the speaker’s: then it may write that voice down instead. When the voice is clear and nobody else is talking, an ordinary speech-to-text tool does as well.
Tell it what you know
Names and topic words. Before reading, type the names and unusual words you expect (the speaker’s name, a place, the subject) under Names and topic words. Where the lips fit one of them, the reading leans towards it. In our tests, typed names and rare words that were really said were misread 20–25% less often, and typed words that were not said were written down in fewer than 1 in 100 chances.
Fix a word. Tap a wrong word, then pick one of its alternatives or type the right one. LipScribe reads the line again with your word in place, so the words next to it can change to fit; the rest of the line stays as it was. Your word is underlined in green, words that changed are highlighted, and every export keeps your corrections. In our tests each correction also put right about 0.3 other words on average, and made about 0.15 others wrong, so do check the highlighted words.
Search the lips for a word
Once a video is read, type a word or a name under Search the lips. LipScribe looks through every line for the stretch where the mouth best fits that word, using what the model already saw (nothing is read again), so it can find a word even where the transcript wrote something else. The matches are listed with their times, best first; tap one to play the video from there.
Each match has a score: how well the mouth fits the word (0 is the best fit). Likely matches are above a line set in our tests so that 9 in 10 of them were the word; weaker ones, below it, were the word less than half the time. On rare words and names that the transcript wrote down only 4% of the time in English (9% in Spanish), the search found 37% of them as likely matches in English and 57% in Spanish, and about 6 and 7 in 10 counting the weaker ones. A match is never proof that someone said a word, and no match is no proof that they did not: in a long video, expect the odd likely match where the word was not said. With Also listen to the sound, the box is hidden: its scores were measured on the lips-only models, and with the sound the transcript itself is far more reliable.
One model per language, downloaded once
Each spoken language has its own model: English 264 MB, Spanish (beta) 169 MB, and lips + sound 286 MB in English, 298 MB in Spanish. Your browser downloads a model the first time you use it and keeps it in its cache, so after that LipScribe starts at once and works offline. Pick the spoken language in the app before reading; it is separate from the interface language.
Everything runs on your own processor. On a laptop CPU a one-minute clip takes about 40 seconds to read; phones and older computers take several times longer.