Real footage, the whole flow — a video goes in, the transcript lands, one line gets fixed, a style gets picked.
Transcription runs on your own free Groq key, and Cap charges nothing on top. Your video never uploads to Cap's servers.
Chinese line-breaking runs on a model Cap trained itself, in your browser through ONNX, with the cloud model demoted to a fallback. Taigi runs on Cap's self-hosted model, the one Whisper doesn't speak.
Text-fix reads the whole transcript, then proposes a replacement table. Every row is homophone-checked, and you can edit or throw out any of them. Nothing touches your subtitles until you press apply.
Cap does almost everything inside your open tab: reading the file, pulling audio, slicing, stitching, line breaks, saving. Only one step ever leaves, and that just sends audio to a transcription service with your own free key.
the one step that leaves
These are the exact seven words on the progress bar. Your video never leaves your machine. Only the audio chunk being transcribed goes out, and only to the key you gave it.
Line breaking used to mean an API call. Now Cap runs its own local model for it, plus a second tiny model that checks if someone is actually talking.
42 MB · Cap's own
We trained this specifically to break subtitle lines where they read naturally. Downloads once, then stays in your browser. If you hit a language it doesn't know, it falls back to a cloud model.
2.3 MB · Silero VAD
A tiny open-source model bundled right with the page. It just answers whether there's a voice in a slice of audio, which keeps the next step safe.
Transcription models love making up polite goodbyes over outro music. Cap spots the usual ones, but won't delete anything unless the voice detector confirms nobody was talking. If it isn't sure, the line stays.
Names, jargon, places. Transcription always trips on the same stuff. Cap reads the finished transcript, works out fixes, and pauses to show you first.
The first pass catches obvious errors. The second runs on the cleaned copy, because some mistakes only show up once the noisy ones are gone.
Every suggestion is checked phonetically against the original. If the model is just trying to reword a sentence, we drop it. Mishearings sound like what was said. Rewrites don't.
Change the wording or ditch the suggestion completely. Nothing in your subtitles moves until you hit apply.
Approved fixes save to your personal list, so Cap already knows them next time.
A tribute to What'Sub — Cap was made because of it.