/video-edit

HyperFrames Transcript Editor

Interactive, browser-only editor for fixing speech-to-text mishears before the final render. Designed for the video-edit skill but works standalone with any transcript.json that follows the faster-whisper word-level format.

What it does

• One-click project folder load — point at a HyperFrames project directory and the editor auto-finds transcript.json + the source/footage video, opens itself, and saves transcript_review.txt back to the same folder when you hit Save (uses the File System Access API; Chrome/Edge 86+, Safari 15.2+; falls back to individual file pickers and a download on unsupported browsers).
• Loads a transcript.json (word-level whisper output) and a source video — entirely in the browser, zero upload.
• Renders each segment as an inline-editable textarea with the timecode prefix.
• Clicking a segment seeks the video to that moment; the active segment auto-highlights as the video plays.
• Auto-saves edits to localStorage so refreshes don't lose work.
• Supports Hebrew RTL (per-segment direction: auto).
• One-click dictionary corrections (paste {wrong: right} JSON, whole-word match).
• One-click find / replace across all segments.
• Optional WebLLM AI suggestions — runs Qwen2.5-3B or Llama-3.2-3B in the browser via WebGPU. First load downloads the model (~1–2 GB, cached forever after). Each segment gets a 🤖 button that asks the LLM to fix Whisper mishears given previous/next-line context, then surfaces an accept/dismiss inline diff.
• Saves a transcript_review.txt in the exact format the skill's apply_review.py expects — drop it back in the project and the agent re-aligns word timings + re-renders.

How to open

# From the video-edit skill directory:
start ./transcript-editor/index.html       # Windows
open ./transcript-editor/index.html        # macOS
xdg-open ./transcript-editor/index.html    # Linux

# Or run a quick HTTP server (recommended for WebGPU/WebLLM)
cd transcript-editor && python -m http.server 8765
# then open http://localhost:8765

WebLLM needs WebGPU support — Chrome / Edge 113+, Safari TP. Firefox stable still lacks WebGPU by default; flip dom.webgpu.enabled in about:config if you must.

Keyboard shortcuts

Workflow with the video-edit skill

1. Agent runs transcribe.py → produces transcript.json.
2. Agent runs make_review.py (applies known corrections, writes transcript_review.txt).
3. Instead of editing the .txt in Notepad, you open this editor, load the transcript.json and source video, fix everything visually with timing + video preview + AI suggestions, and save the new transcript_review.txt.
4. Tell the agent "continue" — it runs apply_review.py to redistribute word timings into transcript.json, regenerates the caption sub-composition, and re-renders.

Standalone use

The editor doesn't need the agent. Drop any whisper JSON + matching video, edit, save. The output is a flat [mm:ss.xx] text file you can copy into your own workflow.

Privacy

Everything runs in your browser. No telemetry. WebLLM model files download from the MLC-AI CDN on first use only; nothing about your transcript leaves your machine.