本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

Rescript

Edit video and audio like you edit text — fully offline, in your browser.

✨ Try it now: wassgha.github.io/rescript

图片：Follow @wassgha on X

图片：Rescript Demo

Rescript is an open-source, transcript-based media editor. Drop in a video or audio file and it is transcribed locally with per-word timestamps and speaker labels. Delete words in the transcript and the corresponding clip is cut from the media. Export the final cut — without your file ever leaving your device.

• 🔒 Private by design — no server, no auth, no uploads; all media processing happens on-device
• 📝 Word-level editing — select words, press ⌫, the cut follows the text
• 🧹 Filler removal — one-click cut of “um”, “uh”, and similar fillers
• 🗣️ Speaker diarization — the transcript is grouped by speaker
• 🎬 Timeline — waveform, word labels, cut regions, playhead, zoom
• ⚡ Live preview — playback skips your cuts in real time
• 📦 In-browser export — frame-accurate MP4 (video) or M4A (audio) with ffmpeg.wasm
• 🎧 Audio files — edit podcasts, voice notes, and interviews the same way as video

Stack

Piece	Tech
App	Next.js + React + TypeScript + Tailwind
Transcription	transformers.js running whisper-base_timestamped or whisper-small_timestamped (WebGPU with WASM fallback) in a Web Worker
Speaker labels	pyannote-segmentation-3.0 (ONNX)
Media processing	ffmpeg.wasm (multi-threaded) for audio extraction and export
State	zustand

Development

npm install     # also copies ffmpeg/onnxruntime WASM into public/vendor
npm run dev     # dev server
npm run build   # production build
npm run lint    # eslint

Open http://localhost:3000 and drop in a video with an audio track.

Note on “offline”: the AI models (Whisper Base ~200 MB, or Small ~600 MB, plus a small speaker model) are downloaded from the Hugging Face Hub the first time you transcribe, then cached in browser storage. After that, everything — transcription, editing, export — works with the network fully disconnected. Your media and transcript never leave the device; the only third-party request the app makes is anonymous page analytics (Google Analytics), which fails silently when offline.

How it works

1. Extract — ffmpeg.wasm decodes the audio track to mono 16 kHz PCM.
2. Transcribe — Whisper runs in a Web Worker with return_timestamps: "word", streaming text as it goes; pyannote assigns a speaker to every word. Choose Whisper Base or Whisper Small on the homepage.
3. Edit — deleting words produces “cut ranges” of the original media. The preview player skips them in real time and the timeline shows them in red. Remove fillers cuts every detected “um” / “uh” / etc. in one click.
4. Export — the kept ranges are trimmed and concatenated with an ffmpeg filter graph and re-encoded (libx264/aac), so cuts are word-accurate.

Browser support

A Chromium-based browser is recommended. The app requires SharedArrayBuffer (served with COOP/COEP headers) and uses WebGPU for inference when available, falling back to WASM otherwise.

License

MIT

---

Built by @wassgha — follow along on X for updates.