Voiseege - Shabbat-Safe Audio Intelligence System

Voiseege is an always-on, Shabbat-safe audio intelligence system designed to run locally on Android devices using Termux. The system operates with zero root access, zero cloud dependency, and strict Shabbat compliance.

Overview

The system is designed to continuously record audio during Shabbat without any semantic processing, and then perform analysis, transcription, and data extraction after Shabbat ends.

System Architecture

The system is divided into three distinct phases:

Phase 1 - Shabbat Mode (Passive Recorder)

• Continuous audio recording in small chunks (15-45s adaptive)
• Store Opus files with timestamps only
• No semantic database writes
• No deletion, dynamic notifications, or CPU upscaling
• Must comply with strict Shabbat rules

Phase 2 - Post-Shabbat Processing

• VAD-based segmentation
• Whisper.cpp transcription (max 120s chunks)
• Detection of Torah Aliyah sales events
• Extraction of buyer name, Aliyah type, amount, and timestamp
• Attach exact audio proof clip to each sale

Phase 3 - Dashboard (Human-in-the-Loop)

• CRUD system for Aliyah sales and congregants
• Audio playback functionality
• Transcript display
• Verification system with confidence scoring
• Human review for low-confidence extractions

Technical Stack

Audio & ML Stack

• Audio capture: termux-microphone-record
• Audio codec: Opus (low bitrate, stable)
• VAD: Silero VAD via ONNX Runtime
• STT: whisper.cpp (compiled locally, ARM64)
• LLM: Local LLM (post-processing only, never during Shabbat)
• Database: SQLite with WAL + atomic write patterns

Hardware & Environment

• Device: Samsung Galaxy S24 Ultra
• CPU: Snapdragon 8 Gen 3 (ARM64)
• OS: Android 14
• Environment: Termux
• Root: Not allowed
• NDK App: Not allowed
• GUI during runtime: Not allowed

Shabbat Compliance Rules

During Shabbat (Absolute Prohibitions):

• ❌ NOT write semantic data (names, money, purchases)
• ❌ NOT classify, infer, tag, or analyze content
• ❌ NOT invoke any LLM
• ❌ NOT allow user interaction
• ❌ NOT modify structured records

During Shabbat (Permitted Actions):

• ✅ Record raw audio automatically
• ✅ Store audio files with timestamps only
• ✅ Run pre-started processes without interaction

All analysis, segmentation, transcription, extraction, and database writing must occur only AFTER Shabbat.

Database Schema

The system implements schemas for:

• raw_audio
• transcripts
• aliyah_sales
• congregants
• verification_log

Each Aliyah sale record links to audio file, transcript, verification state, and human verifier (if any).

Resilience & Recovery

The system implements:

• Supervisor process (never dies)
• Worker isolation (recorder / transcriber / watchdog)
• Memory leak fence (forced restart after N files)
• IO backpressure (token bucket)
• Thermal hard fence (pause above 48°C)
• SQLite WAL + double-write (pending.json)
• File system integrity repair on boot
• Termux wake-lock + foreground notification
• Ultimate keeper loop (restarts supervisor if killed)
• Termux-Boot auto-start after reboot
• Data loss tolerance: ZERO

Aliyah Sale Detection Logic

Detection is rule-based first, LLM second:

• Speech length threshold
• Hebrew keyword patterns (e.g. "נמכר", "עולה", "שלוש מאות")
• Contextual proximity (name + amount)
• Confidence scoring
• LLM is allowed only as a verifier, never as the primary extractor

Installation & Setup

1. Ensure Termux is installed on your Android device
2. Install required packages:

   pkg update && pkg upgrade
   pkg install python nodejs git
   pip install -r requirements.txt

3. Install termux-api for microphone access:

   pkg install termux-api

4. Install whisper.cpp for transcription (optional, for post-Shabbat processing)
5. Configure your Shabbat schedule in config.json

Running the System

Use the provided scripts to manage the system:

Start the system with:

./scripts/start.sh

Stop the system with:

./scripts/stop.sh

View live logs with:

./scripts/live_log.sh

The supervisor will start all necessary components:

• Audio recorder
• Post-Shabbat processor
• Dashboard API

Repository Structure

voiseege/
├── config.json          # System configuration
├── requirements.txt     # Python dependencies
├── audio/              # Audio recordings
├── db/                 # SQLite database
├── logs/               # System logs
├── models/             # ML models (whisper, etc.)
├── scripts/            # Utility scripts
├── src/                # Source code
│   ├── dashboard/      # Dashboard API
│   ├── detector/       # Aliyah sale detection
│   ├── processor/      # Audio processing pipeline
│   ├── recorder/       # Audio recording
│   ├── database.py     # Database management
│   ├── shabbat_manager.py # Shabbat compliance
│   └── supervisor.py   # Main supervisor
└── tests/              # Test suite

Shabbat Compliance Verification

This system has been designed and verified to comply with Shabbat laws:

• No semantic processing during Shabbat
• No user interaction during Shabbat
• No data modification during Shabbat
• All analysis deferred to post-Shabbat

Contributing

Contributions are welcome! Please ensure any changes maintain Shabbat compliance and follow the architectural principles outlined in the documentation.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Repository Setup

To set up this repository for remote pushing:

1. Create a new repository on GitHub or your preferred Git hosting service
2. Use the setup script to add the remote:

   ./scripts/setup_remote.sh <your_repository_url>

3. Push the repository:

   ./scripts/push.sh

Alternatively, you can manually add the remote and push:

git remote add origin <your_repository_url>
git push -u origin main

Support

For support, please open an issue in the repository or contact the maintainers.