Transcript Maker: One Workspace for Captions, Sinhala Translation and SRT Timing

Subtitling a Sinhala YouTube video usually means three tools: one to pull captions, one to translate, and one to fix timing — with copy-paste between them and a lost afternoon when the SRT drifts by half a second.

Transcript Maker puts it in one workspace: paste a YouTube link, correct the text, translate (Sinhala by default), and edit caption timing down to the millisecond, then export SRT or VTT.

The workflow

1. Get transcript. Paste a link; the Flask backend fetches captions with youtube-transcript-api / yt-dlp. Saved transcripts reopen instantly. Videos without accessible captions return a visible error, and Paste transcript instead works with text you already have.
2. Edit. Correct the transcript like a normal document. Paragraphs keep their cue timing; adding or removing paragraphs estimates new timing. Drafts are kept locally; Save changes writes to disk.
3. Translate. Sinhala by default, any language on request. The translation is editable and saved separately from the original.
4. Subtitles / SRT. Edit each caption's words, start and end. Times accept HH:MM:SS.mmm, MM:SS or plain seconds like 12.5.
5. Split here / Merge next / Undo edit for caption surgery. Import an existing .srt or .vtt to edit it.
6. Natural flow. Line-length and duration controls; Break into readable captions redistributes text into smaller cues.
7. Custom ranges. Select captions, enter a start/end, Fit selected captions to range. Download range SRT exports only overlapping cues; Start range SRT at 00:00 re-bases them for a trimmed clip.

These export subtitle files — they don't cut video. The app is honest about that line.

Decisions

Cache everything, deterministically. YouTube rate-limits and IP-blocks repeated caption requests. Every fetched track is cached on disk keyed by a SHA-256 of its source, so the second open of a video is instant and never touches YouTube.

Click-to-seek. HTML5 media events are bound to rendered caption spans: clicking a word seeks playback to that cue. Checking timing becomes clicking, not scrubbing.

Estimated timing is labelled as estimated. Anything the app computes rather than reads — redistributed captions, new paragraphs — is called an estimate in the UI, with a reminder to check against playback.

Focus over features. An earlier version generated voice-overs. It was removed. The app does transcripts and subtitles, and does them well.

Running it

pip install -r requirements.txt
PORT=5050 python3 server.py        # or: docker compose up --build

In Docker it runs under Gunicorn with the transcript cache on a volume, so cached work survives container rebuilds.

See the case study →