Technology Stack

Overview

AI Transcription Notepad combines several key technologies to deliver single-pass voice transcription with AI cleanup.

Core Components

Component	Technology	Purpose
UI Framework	PyQt6	Desktop GUI with tabbed interface, system tray, keyboard shortcuts
Audio Recording	PyAudio	Microphone capture with device selection
Audio Processing	pydub + FFmpeg	Format conversion, compression, gain control
Voice Activity Detection	TEN VAD	Removes silence before API upload to reduce costs
Text-to-Speech	Microsoft Edge TTS	Audio feedback announcements
Database	Mongita	MongoDB-compatible local storage for transcripts
Charts	pyqtgraph	Analytics visualizations
Global Hotkeys	pynput + evdev	System-wide keyboard shortcuts

Transcription Backend

The app sends audio directly to multimodal AI models via OpenRouter for single-pass transcription and cleanup.

Provider	SDK	Endpoint	Notes
OpenRouter	openai (compatible)	openrouter.ai/api/v1	Gemini 3 Flash/Pro, per-key cost tracking

The fused audio pipeline processes recording, AGC, VAD, and compression in a single optimized pass before API submission.

Voice Activity Detection

TEN VAD is a lightweight native library (~306KB) that detects speech segments and removes silence before sending audio to the API. This reduces file size and API costs.

• Bundled with the ten-vad pip package (no download required)
• Sample rate: 16kHz
• Faster and more accurate than Silero VAD for real-time use
• Requires libc++1 on Linux: sudo apt install libc++1

Text-to-Speech

Microsoft Edge TTS powers the voice announcements in TTS audio feedback mode.

• Voice: British English male (en-GB-RyanNeural)
• Pre-generated audio files bundled in app/assets/tts/ (~1.7MB)
• Dynamic generation available for analytics readout
• Uses edge-tts Python package

Python Dependencies

PyQt6>=6.6.0          # Desktop GUI framework
pyaudio>=0.2.14       # Audio recording
openai>=1.40.0        # OpenRouter API (OpenAI-compatible)
pydub>=0.25.1         # Audio processing
ten-vad>=1.0.6        # Voice activity detection
edge-tts>=6.1.0       # Text-to-speech for announcements
mongita>=1.2.0        # MongoDB-compatible local database
markdown>=3.5.0       # Markdown rendering
pynput>=1.7.6         # Keyboard input handling
evdev>=1.6.0          # Linux input device access (hotkeys)
httpx>=0.27.0         # HTTP client
pyqtgraph>=0.13.0     # Charts and visualizations

System Dependencies

Install on Ubuntu/Debian:

sudo apt install python3 python3-venv ffmpeg portaudio19-dev libc++1

Or run the dependency checker:

./scripts/install-deps.sh

File Locations

~/.config/voice-notepad-v3/
├── config.json           # Settings and API keys
├── mongita/              # MongoDB-compatible database
├── usage/                # Daily cost tracking (JSON)
└── audio-archive/        # Opus recordings (if enabled)

Audio Archival

When enabled, recordings are saved in Opus format optimized for speech:

• Bitrate: ~24kbps
• A 1-minute recording uses ~180KB
• Stored in ~/.config/voice-notepad-v3/audio-archive/