Repository navigation
Transcribe
Khryz-Navarro edited this page Nov 9, 2026
·
1 revision
Get the words of a video as:
-
a text transcript (
.txt); -
subtitles (
.srtor.ass) to load in a player; - burned-in captions, drawn onto the video itself.
It works in two steps, from cheapest to heaviest.
-
YouTube's own subtitles, when they exist:
- yt-dlp downloads them with
--write-subs(the creator's) and--write-auto-subs(YouTube's automatic ones), choosing the language and--sub-format srtorass. - This is fast, small and needs nothing new.
- It's a new choice in the download options: "Subtitles: none / English / …", built from the languages yt-dlp reports.
- yt-dlp downloads them with
-
Transcribing locally, for any video or file:
- A speech-to-text engine runs on the computer, such as whisper.cpp.
- Its model is an opt-in download, from about 75 MB (tiny, rough) to about 3 GB (large, best). Settings let you pick, and the size is shown before downloading.
- The output is SRT or VTT with timestamps; a plain transcript drops them.
- Nothing leaves the computer.
-
Burning in:
- ffmpeg's
subtitlesfilter draws the captions on the video. - It needs ffmpeg built with libass. Check the bundled build; ASS styling (font, size, outline) would come from a few presets.
- ffmpeg's
- Should whisper.cpp be bundled in the app or downloaded on first use? Downloading keeps the installer small.
- Should automatic captions be translated into other languages? yt-dlp can fetch YouTube's translated tracks.
- Should the transcript be editable before burning in?
- Speed: local transcription is slow on older computers without a GPU, much slower than the video's length for the larger models. Show progress and time left.
- Size: the models are big, so they must never download without asking.
- Accuracy: varies with language, accents and music. Say it's automatic.
Phase 1: links you paste
Phase 2: tools for your files
Phase 3: new engines
Phase 4: beyond the desktop