Listen - Voice-to-Text App

A multi-platform voice-to-text app with intelligent model routing, allowing you to speak instead of typing.

Features

• 🎙️ System-wide voice recording (Desktop) / One-tap recording (Mobile)
• 🤖 Multiple SOTA STT models with automatic selection:
  • Moonshine (5-15x faster, optimized for edge devices)
  • Distil-Whisper (6x faster, excellent accuracy)
  • Faster-Whisper, Whisper.cpp, Python Whisper
• 🧠 Intelligent model routing - Auto-selects best model for your needs
• 📋 Automatic clipboard copy
• 🪟 Always-on-top overlay (Desktop)
• 📱 Native iOS (Swift + WhisperKit) and Android (Kotlin + TFLite) apps
• 🔒 100% offline - All processing on-device, no cloud services
• ⚡ Ultra-fast transcription

Setup

1. Install dependencies:

npm install

2. Install an STT model (choose one or more):

   🔥 UNDER 1B Parameters (Recommended - Edge-optimized)

   Option A: Parakeet TDT v3 (FASTEST)

   ./install-parakeet.sh  # 600M params, 6.32% WER, 25 languages, ultra-fast inference

   Option B: Moonshine (Mobile-optimized)

   ./install-moonshine.sh  # 40-200M params, 5-15x real-time

   Option C: Distil-Whisper (Best for English)

   ./install-distil-whisper.sh  # 244M params, 6x real-time

   Option D: Faster-Whisper (Good balance)

   pip install faster-whisper  # 74M params, 4x real-time

   Option E: whisper.cpp (C++ implementation)

   ./setup-whisper.sh  # 74M params, 2x real-time

   Option F: Python Whisper (Fallback)

   ./install-python-whisper.sh  # 74M params, baseline

   🎯 OVER 1B Parameters (Optional - Maximum accuracy)

   Option G: Canary Qwen 2.5B (#1 Accuracy)

   ./install-canary.sh  # 2.5B params, 5.63% WER, 

   Note: The app will automatically use the fastest available model. Install multiple models for automatic fallback. Only models you install will be used.

3. Build and run:

npm run build
npm start

Or run in development mode:

npm run dev

Usage

1. Press Ctrl+Shift+Space to activate the overlay
2. Speak your text
3. Press Ctrl+Shift+Space again to stop recording
4. The transcribed text will be automatically copied to clipboard
5. Paste (Ctrl+V) in any application

Keyboard Shortcuts

• Ctrl+Shift+Space - Start/Stop recording
• Esc - Cancel recording and close overlay

Model Selection & Routing

Listen uses an intelligent routing system that automatically selects the best available model based on your requirements.

Recommended Models:

• Desktop (English): Distil-Whisper Small (6x faster, excellent accuracy)
• Desktop (Multilingual): Moonshine Base (5-15x faster, good accuracy)
• Mobile (iOS/Android): Moonshine Tiny (ultra-fast, only 40MB)

See MODEL_COMPARISON.md for detailed benchmarks and comparisons.

Platform Support

• ✅ Linux (Desktop - Electron)
• ✅ iOS 16+ (Native Swift app) - See mobile/ios/README.md
• ✅ Android 7+ (Native Kotlin app) - See mobile/android/README.md
• 🔜 macOS (Desktop - Coming soon)
• ✅ Windows (Desktop - Initial support)

Project Structure

listen/
├── src/                    # TypeScript source code
│   ├── models/            # STT model implementations
│   ├── assets/            # UI (HTML/CSS)
│   └── main.ts            # Electron entry point
├── scripts/               # Python utility scripts
│   └── record_audio_windows.py
├── docs/                  # Documentation
└── mobile/                # Native iOS & Android apps

See ARCHITECTURE.md for complete structure.

Documentation

• Architecture Overview - System design and modular architecture
• Model Comparison - Detailed STT model benchmarks
• Quick Start Guide - Get up and running in 5 minutes
• iOS README - iOS app documentation
• Android README - Android app documentation

Requirements

• Node.js 18+
• One of: whisper.cpp, Python whisper, or faster-whisper (local models, no API needed)
• Audio recording: arecord (ALSA) or sox on Linux

License

MIT