Practical notes on subtitle QC, AI transcription, line-breaking, and localization workflows — from the team building GeekLink.
Why AI-translated subtitles can develop gaps or shifted timing — and how original SRT timecodes, stable IDs, and structured JSON keep every line aligned.
Read more → Product UpdateTypewriter and fade-in captions reveal text a character at a time. GeekLink 0.4.34 rebuilds those temporary OCR fragments into cleaner, editable subtitle cues.
Read more → AI AccuracyOpen-source 1,140-sample, 6-language OCR benchmark: PaddleOCR, EasyOCR, Tesseract, and GeekLink's own models all lose ~2.6-2.9x accuracy near a watermark — then we validated the ranking against real production data, and it flipped.
Read more → Product Update0.4.4 first brought OCR to Windows. 0.4.5 completes the Windows build with speech recognition, OCR, AI translation, editing, batch export, and burn-in.
Read more → Product UpdateNo more guessing where the subtitles are. GeekLink's OCR now scans the video, detects each subtitle line, and pre-fills the language — you just confirm. And with 0.4, GeekLink is out of beta.
Read more → AI WorkflowChatGPT and Claude can't reliably transcribe a batch of videos. Here's the workflow I use: batch-convert videos to text with speech recognition and OCR, then feed clean transcripts to AI to summarize, quiz, and answer questions.
Read more → AI AccuracyI ran GeekLink's speech recognition on a TED talk, then checked every line against TED's own official captions — here's exactly where timing and wording differed.
Read more → Subtitle EditThe Subtitle Edit features that save the most time when reviewing subtitles — bookmarks, Multiple Replace, Batch Convert, and more.
Read more →