OCR your hardsubs without calling a single API

What it does

Video-subtitle-extractor (VSE) is a GUI and CLI tool that pulls hard-coded subtitles out of video files and spits out SRT or plain text. It handles the full pipeline: keyframe extraction, text region detection, OCR, deduplication, and filtering out watermarks or station logos. It also pairs with a separate “subtitle remover” project if you want to erase the original hardsubs entirely.

The interesting bit

The whole thing runs locally on PaddlePaddle models — no Baidu, Alibaba, or any cloud OCR keys to beg for. The README is refreshingly honest about the trade-offs: “fast” mode may drop a few lines and typo here and there; “accurate” mode is frame-by-frame, nearly perfect, and “very slow.” The sweet spot is “auto,” which picks model size based on whether it detects a GPU.

Key highlights

• Supports 87 languages including Chinese (simplified/traditional), English, Japanese, Korean, Arabic, and most major European languages
• Three speed/accuracy tiers: fast, auto (recommended), and accurate
• GPU acceleration via CUDA, DirectML, or ONNX; CPU fallback available
• Batch processing for multiple videos with matching resolution and subtitle region
• Custom text replacement/deletion via typoMap.json — useful for cleaning up watermarks or common OCR errors
• Cross-platform: Windows, macOS, Linux

Caveats

• Paths containing Chinese characters or spaces will “cause unknown errors” — the README warns this explicitly and repeatedly
• NVIDIA 50-series cards need CUDA 12.8+, but Paddle 3.3.1 doesn’t support it yet; DirectML workaround suggested
• ONNX path for macOS/AMD is marked “untested” and the maintainers ask you not to file issues about it

Verdict

Worth a look if you regularly need to archive, translate, or repurpose subtitled video and would rather not pipe your content through someone else’s API. Skip it if you need real-time processing or can’t tolerate the path-naming restrictions.