Escriba

A local browser-based audio transcription tool that runs entirely in your browser using Whisper (ONNX) with WebGPU or WASM acceleration.

Features

• 100% Local Processing - All transcription happens in your browser, no server uploads
• GPU Acceleration - Uses WebGPU for faster transcription when available, with automatic WASM fallback
• Voice Activity Detection - Smart silence detection for optimal audio segmentation
• Multi-Worker Processing - Parallel transcription using up to 3 workers for faster processing
• Text Post-Processing - Automatically removes repeated sentences, words, and bracketed content
• Multiple Model Sizes - Choose between Tiny, Base, Small, or Medium (English) Whisper models
• Language Selection - Support for Portuguese, English, Spanish, French, German, Italian, Japanese, Chinese, Korean, plus auto-detect
• Batch Processing - Transcribe multiple audio files at once
• Audio Recording - Record directly from microphone and transcribe instantly
• Export Options - Download individual text files, ZIP of all transcriptions, or copy to clipboard
• Performance Stats - View processing time per file and total
• Cross-Platform - Works in any modern browser with WebGPU or WebAssembly support

Usage

1. Open the page in a browser (Chrome, Edge, or other WebGPU-enabled browsers recommended)
2. Select your preferred Whisper model (Tiny, Base, Small, or Medium.en)
3. Choose your language or use auto-detect
4. Upload audio files or record directly from microphone
5. Click "Transcrever" to start transcription
6. Copy or download results

Live Demo

Visit: https://afa7789.github.io/Escriba/

Available Models

Model	Size	Languages	Default
Tiny	39MB	Multi
Base	74MB	Multi	Yes
Small	241MB	Multi
Medium.en	769MB	English only

Supported Languages

Portuguese, English, Spanish, French, German, Italian, Japanese, Chinese, Korean, or auto-detect.

Tech Stack

• Whisper ONNX - Speech recognition models
• Transformers.js - Running Hugging Face models in the browser
• WebGPU / WebAssembly - Hardware acceleration
• Vanilla JavaScript - No build step required

Development

npm install     # Install dependencies
npm run serve  # Start local server
npm run lint   # Run ESLint
npm test       # Run tests

Or use the Makefile:

make install   # Install dependencies
make serve     # Start local server
make lint      # Run ESLint
make lint-fix  # Auto-fix lint issues
make test      # Run tests

Project Structure

Escriba/
├── index.html        # Main HTML page
├── styles.css        # CSS with light/dark theme
├── worker.js         # Web Worker for Whisper transcription
├── package.json      # Dependencies and scripts
├── src/
│   ├── app.js        # Main application logic
│   ├── state.js      # Global state management
│   ├── transcriber.js # Transcription orchestration
│   ├── audio.js      # Audio recording & preprocessing
│   ├── files.js      # File handling & UI
│   ├── device.js     # WebGPU/WASM detection
│   ├── dom.js        # DOM utilities
│   ├── results.js    # Results display & download
│   ├── vad.js        # Voice Activity Detection
│   └── merger.js     # Segment merging
└── tests/            # Test files