01 // Current Command Processing Flow
The current implementation does not use a three-lane command router. The executor handles pending security confirmations and greeting requests directly, routes explicit GitHub requests to the built-in GitHub plugin, and sends other non-empty requests to the Agent Brain. The Agent Brain asks LiteLLM for Python code using the configured model identifier ollama/gemma4:31b-cloud, then runs returned code in a Python subprocess.
The model identifier includes cloud; do not assume requests are processed offline or that prompt text remains on the device. Generated code is executed with the application's user permissions. The website command simulator is illustrative and does not submit commands to the running application.
02 // Speech & VAD Pipeline Upgrades
Continuous audio capture operates via a high-performance 16kHz circular buffer managed by sounddevice and filtered in real time by Silero Neural VAD:
- Phonetic Wake Word Expansion: The engine automatically matches phonetic variations including
"Privacy 68","Privacy sixty eight","Privacy sixty-eight","Privacy six eight", and digit-separated aliases. - 2-Gram Multi-Word Fuzzy Matching: Applies sliding Levenshtein distance across 2-word sequences, accommodating fast speech, regional pronunciation, and low-fidelity headsets.
- Dynamic Peak Normalization: Every audio frame is normalized against its peak amplitude before transcription, ensuring quiet voices trigger as reliably as studio microphones.
- Silero VAD Gating: 80ms audio frames are analyzed by the Silero neural network. Non-speech background noise is rejected, dropping CPU idle usage to under 1.5%.
- Whisper Decoder Priming: Decoder token probabilities are primed with domain hotwords via
config.INITIAL_PROMPTto prevent dropping numerical identifiers.
03 // Microphone Setup & Audio Diagnostics
Proper microphone levels ensure 100% accurate wake word recognition. Follow these diagnostic steps:
1. Windows Privacy Permissions
Ensure Windows allows desktop apps access to the microphone: Navigate to Settings → Privacy & Security → Microphone. Toggle "Allow desktop apps to access your microphone" to ON.
2. Audio Level Benchmark
Test your microphone's live input sensitivity and background noise floor by running this test command in PowerShell:
python -c "import sounddevice as sd, numpy as np; d=sd.rec(int(16000*3), 16000, 1, 'float32'); sd.wait(); print('Peak:', round(float(np.max(np.abs(d))), 4), 'RMS:', round(float(np.sqrt(np.mean(d**2))), 4))"
| ACOUSTIC STATE | ACCEPTABLE RMS | ACCEPTABLE PEAK | RECOMMENDED ACTION |
|---|---|---|---|
| Room Silence | < 0.005 |
< 0.015 |
Optimal noise floor. VAD will remain in low-power idle scan. |
| Normal Speech | > 0.020 |
> 0.080 |
Ideal speech level. Whisper transcribes with 99%+ accuracy. |
| Low Sensitivity | < 0.010 |
< 0.030 |
Increase microphone volume to 85% in Windows Sound Control Panel. |
04 // Biometric Voice Authentication (Voice Lock)
PRIVACY68 features biometric speaker authorization powered by ECAPA-TDNN / CAM++ neural embeddings (512-dimensional vector space). When activated, commands are only executed if vocal tract biometrics match the enrolled master profile.
Enrollment Procedure
- Open the PRIVACY68 Control Center (Right-click system tray icon → Open Dashboard).
- Navigate to the Voice Profile / Biometrics tab.
- Click "Enroll Master Voice" and recite the 3 calibration phrases clearly.
- The model extracts normalized 512-d acoustic embeddings, averages the vector cluster, and stores it in
%APPDATA%/PRIVACY68/voice_lock/.
Threshold Tuning
| THRESHOLD | SECURITY PROFILE | BEHAVIOR & ENVIRONMENT |
|---|---|---|
0.40 – 0.45 |
Relaxed | Accommodates head colds, background music, or varying distances from mic. |
0.50 (Default) |
Balanced | Rejects unfamiliar speakers while reliably recognizing the owner. |
0.65 – 0.75 |
Strict Biometric | High security mode. Rejects slight vocal variations and voice recordings. |
Note: Adjusting the threshold in the dashboard takes effect dynamically in memory with zero app restarts.
05 // Hardware Acceleration & Performance Boost
PRIVACY68 automatically detects your host hardware on startup via utils/cuda_manager.py and configures optimal compute precision:
| HARDWARE SETUP | COMPUTE PRECISION | TRANSCRIPTION TIME | MEMORY USAGE |
|---|---|---|---|
| NVIDIA GPU (CUDA 12) | float16 (Tensor Cores) |
~120ms – 160ms |
~1.4 GB VRAM |
| Standard Intel / AMD CPU | int8 (AVX2 / AVX-512) |
~350ms – 600ms |
~1.5 GB RAM |
Zero-Config CUDA Runtime Manager
For users with NVIDIA GPUs who do not have the complete CUDA toolkit installed, the built-in CUDA Manager allows one-click runtime installation. Simply click "Download & Register CUDA 12 Runtime" in the Dashboard to download the isolated cuBLAS and cuDNN DLLs to your user profile.
06 // Plugin Architecture & Mobile Web Remote
The desktop dashboard includes a plugin registry for viewing registered integrations, changing enabled state, creating a template, installing a Python plugin file, and uninstalling user plugins. It also supports connecting a GitHub personal access token through Windows Credential Manager. External Python plugins are loaded from %APPDATA%/PRIVACY68/plugins/; plugin code runs with the current user's permissions.
Explicit GitHub requests are routed directly to the built-in GitHub integration for current-repository status/push, repository and issue/pull-request listings, and issue/pull-request creation. Pushes and create actions require confirmation. Other plugin registrations are not automatically dispatched. The plugin registry is not a sandbox or an execution boundary.
See the built-in plugins and extension details.
Mobile Web Remote (Local Wi-Fi)
Control your computer from your smartphone or tablet over the local network. The remote server does not provide authentication; only enable it on a trusted network and keep in mind commands run on the host computer.
07 // Complete Voice Command Reference
These examples describe current dispatch behavior. GitHub requests listed below use the direct integration; most other requests go to the Agent Brain, and outcomes depend on model response, available software, and permissions.
| CATEGORY | EXAMPLE VOICE COMMANDS | CURRENT HANDLING | NOTES |
|---|---|---|---|
| System Control | "Privacy68, mute volume", "set volume 75" |
Agent Brain | Requests are sent to the configured model; actions are not guaranteed. |
| Display & Power | "Privacy68, lock screen", "take screenshot" |
Agent Brain | Generated code runs in a subprocess with the current user's permissions. |
| Applications | "Privacy68, open Chrome", "launch VS Code" |
Agent Brain | Results depend on model output and installed applications. |
| Web Navigation | "Privacy68, open YouTube", "open GitHub" |
Agent Brain | Requests may involve network access if generated code performs it. |
| Windows Settings | "Privacy68, open bluetooth", "open wifi settings" |
Agent Brain | No separate deterministic settings lane is implemented. |
| Compound Browser | "Privacy68, play Believer on YouTube" |
Agent Brain | The website demo does not execute this example. |
| Messaging | "Privacy68, send WhatsApp message to Mom" |
Agent Brain | The built-in plugin code is available, but the executor does not route commands directly to it. |
| Live Web Search | "Privacy68, search latest space discoveries" |
Agent Brain | Any live search depends on model-generated code and may use the network. |
| General AI Reasoning | "Privacy68, explain quantum computing simply" |
Agent Brain | Current model identifier: ollama/gemma4:31b-cloud. |
| GitHub | "check GitHub status", "push my code to GitHub branch feat/demo", "list repositories", "show issues" |
GitHub plugin | Connect a fine-grained token in the dashboard. Pushes and GitHub write actions require confirmation; pushes do not stage or commit files. |
08 // Configuration Reference (`config.py`)
Core thresholds, model selections, and device parameters can be fine-tuned in config.py: