Scriber

A full-screen Markdown editor with voice dictation. Speak, pause, and the recognized text is appended as a new paragraph. The editor is a CodeMirror Markdown editor (list continuation on Enter, syntax highlighting, follows the browser light/dark theme).

This repo is only the frontend (deployed to GitHub Pages). It is a plain OpenAI-compatible client and does no ML itself — it talks to a backend that exposes /v1/audio/transcriptions (e.g. rapid-mlx serving NVIDIA Parakeet).

How it works

• Microphone → AudioWorklet captures PCM16 @ 16 kHz (public/pcm-worklet.js).
• A browser VAD (energy-based) segments speech; on a pause the utterance is encoded to WAV and POSTed to /v1/audio/transcriptions.
• The returned text is appended to the document at the end (the caret is preserved, so you can keep editing elsewhere while dictating).
• Ctrl/Cmd+Enter commits the current utterance immediately.
• Stopping the recording auto-copies the whole document to the clipboard.

Endpoints are same-origin /v1/* by default, so a local proxy serves this page and routes /v1 to the backend. A hosted service can be set in Settings (Backend URL + API key).

Run locally (dev)

npm install
npm run dev            # http://localhost:5173

The dev server proxies /v1 to a local backend by sub-path (see vite.config.ts): /v1/audio → :8002, /v1/chat → :8000. Point those at your local OpenAI-compatible servers (e.g. the local-llm rapid-mlx stack running Parakeet on :8002).

The mic needs a secure context: http://localhost works directly, or serve over HTTPS (e.g. tailscale serve) when opening from another device.

Build

cd front
npm run build         # static output in front/dist/

Related

• Backend: any OpenAI-compatible server exposing /v1/audio/transcriptions (Parakeet via rapid-mlx, a Whisper server, etc.).
• Local proxy (scriber-local, separate repo): serves this frontend and routes /v1 to the local backend so everything is same-origin on localhost.