Chrome extension · Open beta · Free
Stop reading the bottom of the screen. Watch the face. EyeStill puts every word right at the speaker's lips, the moment it is said — you read and see the face in one glance. It follows whoever is talking and moves with every cut. Watch at 1× or 2×: your eyes stay where the person is. Turn on translation and keep the real voice, while the meaning appears at the lips, translated with the whole talk in mind. ChatGPT, Claude and Gemini answers arrive the same way, while they are still being written. Books too. One point of focus, no lines to chase. Free, no account, and the video never leaves your browser. You've just read all this without moving your eyes.
Captions sit at the bottom of the frame; the person — expression, lips, emphasis — is in the middle. Your eyes bounce between the two several times a sentence, more at 2×, and every trip drops a little of the words or a little of the face. AI answers do the same on a page: five paragraphs land at once, and you skim for the point.
EyeStill finds the speaker and puts the subtitle line just above their lips, a word at a time, as it is spoken. When two people talk, it follows whoever is speaking; when the shot cuts, it moves with it.
YouTube · word-timed captions · translation
ChatGPT, Claude, Gemini, DeepSeek. A button under each answer reads it to you right there, word by word at one point of focus — no scrolling, no hunting for the paragraph you were on.
Alt + Q in any conversation
More about ai answers →Open an EPUB, FB2 or TXT in the EyeStill reader, in any browser. Your library and your place are kept in the browser. A 90,000-word novel becomes a few evenings instead of a few months.
EPUB · FB2 · TXT · pasted text
More about books →Open an article, a public-domain book or Claude's own answer in the reader inside Claude. Pause on a hard sentence and ask about it — Claude knows which sentence you stopped on, and can recap, quiz you or pick up where you left off.
Claude connector · /eyestill plugin
How it works →Fast speech blurs and attention drifts to the phone. Each word lands where you're already looking — on the speaker — so 2× stays readable and the face keeps you in.
You follow English until a podcast guest speeds up or slips into slang. Watch the lips and read the exact word the moment it is said.
Dubbing swaps the voice; bottom subtitles turn a talk into reading. Keep the speaker's own voice and expression, and read the translation where they're speaking.
Placing a word at someone's lips sounds simple. Doing it on any video — through cuts, gestures and two-person shots, in the speaker's rhythm and in your language — takes vision models, a speech pipeline and a translator working together. The vision part runs entirely in your browser: the models ship inside the extension and use your GPU when there is one.
MediaPipe BlazeFace looks at a small copy of the frame and returns the mouth point. In wide shots it re-scans the upper frame as two zoomed halves, and a body-pose model finds people turned to each other that a face detector misses.
The line's bottom edge rests just above the upper lip, so the mouth stays visible, and the red focus letter is centred over it. Letterbox bars are measured; vertical Shorts and 4:3 video are not squashed.
Eight times a second it compares how much each mouth changed with how much the eyes of the same face did. Speech moves the mouth, a nod moves both. A new speaker must be 1.8× more active, twice in a row, to take the line — a listener's “uh-huh” doesn't.
A scene change is caught from the median colour change over an 8×8 grid of the frame, so the line jumps to the new face at once. A gesturing hand changes only a few cells and is ignored.
A median over the last five looks, a 70-pixel dead zone and three confirmations before a long jump: nods, gestures and a hand across the face don't move it. When the face leaves for slides or b-roll, the line waits a few seconds, then glides down.
Corner webcams in screencasts, small avatars and faces at the edge of the frame are filtered out, so a tutorial's subtitles stay at the bottom where they belong.
YouTube's auto-captions carry a time for each word. EyeStill reads the same caption file the player downloaded, so each word appears the moment it is said. Line-timed captions are spread over the line by length.
YouTube's rolling caption lines repeat words at every seam. A seam matcher stitches them into one stream, so nothing is read twice — while a real “we're, we're” in the speech survives.
English auto-captions come without punctuation. Chrome's built-in language model punctuates them ahead of playback, and a chunk is used only if the model kept every word.
A large language model translates the transcript about a minute and a half ahead, with the neighbouring lines as context: the same term stays the same term, phrasing sounds spoken, and fillers like “you know” and false starts are dropped.
Each translated line takes its original's place and time. The first lines around where you are come back in about a second; a video translated once is cached, so the next viewer usually gets it at once. Without our server, Chrome's on-device translator steps in.
Word-frequency data for 30,000 words per language and a surprisal model of what usually comes next decide how long each word stays: rare and surprising words a beat longer, predictable ones less, at the same average speed.
Reading a page is a sequence of small leaps — about four a second — each landing on a new cluster of letters, each a chance to drift, re-read or lose the line. A good part of reading time goes into moving the eye rather than understanding.
EyeStill takes that part away. Every word arrives at the same point, shifted so that the letter your eye recognises first sits exactly in the centre — the red one.
No jumps between lines, no lost place, no re-reading loops. The eye stays still; the words do the moving.
On video that point sits at the speaker's lips, so expression and articulation stay in view while you read what they say.
Each word appears as it is spoken. Rare words get a little longer, frequent ones less, commas breathe — the rhythm follows the language.
An interview, an answer, a chapter — read it where your eyes already are.