AI-powered desktop tool for correcting and polishing ASR-generated subtitle files using knowledge bases.
Built for academic lecture videos with mixed Chinese-English content. Import subtitle files from Adobe Premiere or other tools, correct them with AI + domain knowledge, export high-quality subtitle files.
Auto-generated subtitles for Chinese academic lectures are full of errors — misspelled terms, wrong proper nouns, garbled English vocabulary, filler words, and ASR hallucinations. Manual correction of a 60-minute lecture takes hours. Professional subtitle services cost hundreds of dollars per video.
Caption KB Assistant uses AI (Claude API / OpenAI-compatible API) combined with your own knowledge base (lecture slides, posters, terminology lists) to automatically correct subtitle text. A 60-minute lecture costs ~$0.30–0.50 in API fees and takes 3–5 minutes to process.
- SRT Import/Export: Import subtitle files from Adobe Premiere, Subtitle Edit, CapCut, or any SRT-compatible tool
- Plain Text Extraction: Extract text from SRT for editing or external processing
- Two-Tier Knowledge Base:
- Global KB: Personal profile, institution info, research area — reused across all projects
- Project KB: Lecture-specific materials (PDF, PPTX, DOCX, TXT) — per video
- Terminology Table: Maintain a growing list of ASR error → correct term mappings; works offline without API
- AI Correction with Two Modes:
- Basic: Fix typos, correct terminology and proper nouns from knowledge base only
- Advanced: Full polish — remove filler words, rewrite for fluency, improve logical flow
- Automatic Chunking: Handles long lectures (27,000+ characters) by splitting into batches with context overlap
- Diff View: Side-by-side comparison of original vs. corrected text with accept/reject per entry
- Timing Offset Adjustment: Fix systematic time drift in exported SRT files
- Subtitle Styling: Basic font, size, color, and border settings (Chinese: SimSun/SimHei; English: Times New Roman/Arial)
- Multiple Export Formats: SRT, ASS (with styling), TXT, Markdown correction report
- Multi-LLM Support: Claude API (primary), OpenAI-compatible API (secondary)
- Fully Local Data: All files stored locally — nothing uploaded to cloud except API calls
- Bilingual UI: English / 简体中文
Adobe Premiere (generate ASR subtitles)
↓ Export SRT
Caption KB Assistant (AI-powered correction)
↓ Export corrected SRT
Adobe Premiere (import corrected SRT, adjust styling, export video)
Step 1: Import → Load SRT file, view subtitle list, adjust timing
Step 2: Extract → Extract plain text for review and editing
Step 3: Knowledge → Upload lecture materials (PDF, PPTX, DOCX)
Step 4: Correct → AI correction with cost preview, diff view, accept/reject
Step 5: Merge → Combine corrected text with original timestamps
Step 6: Export → Download SRT/ASS/TXT files
Download Caption KB Assistant Setup x.x.x.exe from Releases. Run the installer and follow the prompts.
Download Caption KB Assistant x.x.x.exe from Releases. No installation needed — just run the exe.
- Open the app and go to Settings
- Set your data folder — where project files will be stored
- Enter your API key — Claude API key (get one from Anthropic Console)
- Go to Global KB — fill in your personal profile (name, institution, research area)
- Create a new project — import an SRT file from Adobe Premiere
- Upload knowledge base — add lecture poster, slides, or scripts
- Run AI correction — choose Basic or Advanced mode
- Review the diff — accept, reject, or edit changes
- Merge and export — download the corrected SRT file
- Import back into Premiere — File → Import → select the corrected SRT
- Platform: Electron (Windows)
- Frontend: React + TypeScript + Vite + Tailwind CSS
- AI: Claude API, OpenAI-compatible API
- Storage: Local files (YAML, Markdown, SRT, TXT)
- Build: electron-builder (NSIS + portable)
# Install dependencies
npm install
# Run in development mode
npm run dev
# Type check
npx tsc --noEmit
# Build for production
npm run build
# Package for Windows
npx electron-builder --win --x64This project grew out of a real need: producing Chinese subtitles for academic lecture videos at Arizona State University. After testing five different auto-captioning tools (Camtasia, Adobe Premiere, CapCut English, CapCut Chinese, Subtitle Edit + Whisper) on the same 60-minute lecture, we found that:
- No single tool produced usable Chinese subtitles out of the box
- English academic terms mixed into Chinese speech were consistently garbled
- Proper nouns (researcher names, university names) were almost always wrong
- AI correction using knowledge base context (lecture slides, posters) dramatically improved accuracy
This tool automates the correction workflow that was previously done manually in Claude's chat interface.
- Email KB Assistant — AI-powered email drafting tool with knowledge base, built with the same tech stack