Show HN: Overtone – Transcribe, correct and cut captions for any screen shape

Overtone is an open-source toolkit for creating captions that adapt to different screen shapes. Instead of storing only finished subtitle cues, its .overtone.json format records each spoken word once, with timing and speaker information, then stores separate caption layouts for formats such as widescreen and vertical video. Its CLI can recut existing SRT or WebVTT subtitles, transcribe audio or video, and export standard subtitle files; a video.js plugin displays the layout suited to the player. Layout rules aim to keep names and numbers together, avoid ending lines on weak words, separate speakers, and use pauses as breaks. Recutting works offline, while transcription can use cloud speech services or local tools.

For transcription, several speech engines supply words, timings, and speaker labels; disagreements are reviewed by text models, and an LLM can correct terms, mark emphasis, and flag uncertain edits for human review. It does not assign timestamps. In a small benchmark, Overtone’s English line breaks ended on weak words 0.3% of the time, compared with 28.8% for length-based wrapping and 16.1% for an ElevenLabs export; it also marked all 47 measured pauses, versus 28% and 72% for those alternatives. Across five clips where all engines produced text, its word error rate was 5.3%, better than either individual engine. The project supports full line-cutting rules for English and Dutch, has more limited handling for several other languages, and does not provide translation, sound-effect captions, or a graphical editor.