Open-source captions · MIT
One file holds the words and a line cut for each shape: 16:9, 9:16 or any you add. The player shows the cut closest to the picture. Re-cut subtitles you already have, or make new ones from audio.
Most captions are cut once, for one screen, by filling each line to a length limit. Turn the video vertical and that cut falls apart. Overtone stores every word once, with its time and speaker, and cuts the lines for each screen shape by reading rules. A line avoids ending on “the” or “of”. Names and numbers stay together. Each speaker gets their own caption, and pauses show as breaks. Bring your own SRT or WebVTT, or let the CLI transcribe your media: a transcription model writes the text, speech engines time each word, and an LLM editor fixes names and terms. The LLM never sets timestamps.
A NASA interview with two speakers, cut for a 9:16 phone video (22 characters × 2 lines), from the same words and times. Filling lines to the limit ignores who is speaking and ends lines on “Jim,”, “A”, “I”, “the” and “about”.
the 2004 class. Jim, thank you for joining us. It's great to be here, Allard. Thanks. Well, if you wouldn't mind, we've got a … don't you walk us through you? Okay. A little about me: I grew up in Eugene, Oregon. I joined the Astronaut Corps about four years ago, as you said, May of 2004,
Jim, thank you for joining us. It's great to be here, Allard. Thanks. Well, if you wouldn't mind, … Okay. A little about me: I grew up in Eugene, Oregon. I joined the Astronaut Corps about four years ago, as you said, May of 2004,
On three English clips (21 minutes, 16:9 cuts), 0.3% of Overtone’s breaks end on a word like “the”, “of” or “and” with no pause after it. Filling lines to the limit gives 29%, ElevenLabs’ SRT export 16%. Overtone shows all 47 pauses of 0.5–1.5 s as a break; the other two show 28% and 72%. These are rule checks, not reader judgments. Details in the caption benchmark.
The format sits in the middle. The player shows it, and the CLI writes it.
overtone-formatOne .overtone.json file holds each word once, with its time, speaker and semantic classes, plus a line cut per screen shape. Comes with a JSON Schema, a validator and exporters to WebVTT, SRT and plain text. No runtime dependencies. Runs in Node and the browser.
{ "text": "Peacocks.",
"start": 30.39, "end": 31.07,
"classes": ["key"] }videojs-overtoneA video.js 8.5+ plugin. It shows the line cut closest to the shape of the picture and switches live on resize. The CC button or the C key cycles off → styled → plain. Pick one of 5 style packs or write your own CSS.
player.overtone({
src: 'talk.overtone.json',
theme: 'social',
});overtone-captionsRe-cuts the SRT or WebVTT you already have, offline and with no keys. Or turns audio and video into captions: text from a transcription model (OpenAI, Azure OpenAI or ElevenLabs), word times and speakers from Azure AI Speech (or ElevenLabs or OpenAI), and an LLM editor (OpenAI, Azure OpenAI or Anthropic). Code checks the structure of every edit. --local runs it all on your machine. Writes JSON, WebVTT, SRT and a transcript, or an MP4 with the styled captions burned in.
npx overtone-captions recut talk.srt
overtone transcribe talk.mp4 \
--context "Panel about cycling"Everything here runs in your browser from static files. The videos stream from NASA and Blender; they are not part of this project.
Three Blender open movies and a NASA interview in a 16:9 and a 9:16 player. Switch style packs, show or hide speaker names and cycle the CC button.
Open the demo →Every semantic class in all 5 style packs, styled and plain, at 16:9 and 9:16, with a check that each line fits.
All packs side by side →Unedited files from a real pipeline run on the NASA interview: 473 words, two speakers.
The same for three Blender open movies, each playable in the demo:
Each task is small, has its own guide and needs no knowledge of the pipeline.
A language pack is a set of word lists written by native speakers. They tell the line cutter which words belong together and where a line may end.
Language guide →A caption theme for the player: one CSS file and a short pack.json. No JavaScript. The tests tell you when it is ready.
Connect another transcription, timing or LLM service to the CLI, through the interface every provider uses.
Provider guide →