Overtone open-sources a multi-model caption tool that transcribes, corrects and reformats subtitles

Lont co-founder Ruud has released Overtone, an MIT-licensed open-source tool for creating captions for video. It sends audio to several speech-to-text models, and when their outputs conflict, re-checks small clips and has an LLM choose the most plausible wording. A rules-based cutter then splits transcripts into captions suited to different screen shapes, such as 16:9 and 9:16, grouping lines by speaker and keeping names and numbers intact. The project also includes a videojs plugin for toggling captions and a command to burn them into MP4 files. The creator says the approach lowers word error rates on test files but costs more than single-model transcription, so it may not suit very high-volume uses.