VobSub to SRT Tool

• Privacy: your files are held in the server's volatile memory only while the job runs and are never written to disk. That much is necessary: unknown glyphs are shown to a vision model, and their shapes are added to the shared glyph memory, which improves the tool for everyone; newly learned glyph sets are published to the GitHub repository regularly. No words or text are kept - the finished SRT is stored in your browser.
• Known glyph set: a few seconds, no vision model involved. This is the case for every glyph set that the tool learned before (currently ? glyph sets in memory).
• New glyph set: the converter asks a vision model about the cues that teach the most new glyphs, learns their shapes, and reads the rest itself. Typically 50-200 requests, a couple of minutes; each request is made once per glyph shape, never per cue.
• Accuracy: glyphs are matched pixel-exactly, so text is reproduced literally (typos included). Where the vision model and the memory disagree, the memory wins once it has enough evidence. Pixel-identical I/l are decided by dictionaries.
• Unknown glyphs that could not be learned (budget used up, no vision model) are written as � so you can find them.
• Queue: add as many tracks as you like; they are converted one after another. Jobs and their SRTs stay in this browser's storage until you delete them, so you can download them again any time.
• Limits per user: ?.
• Need more? The tool is open source. Learn how it works and host it yourself, with your own vision model and no limits at all: full README on GitHub.

Every glyph set the converter has learned so far, newest first. Click one to review its glyphs: the bitmap, how it is read, how often it was seen.