Detect AI-cloned voices. Protect your identity. Everything runs locally — your voice data never leaves your machine.
- Register your voice → creates a unique voiceprint
- Verify a suspicious audio → tells you if it's really you or a deepfake
- Scan any audio → checks for AI-generated artifacts
No cloud. No accounts. No data collection.
You need Node.js 18+, Python 3.9+ and ffmpeg.
git clone https://github.com/rankgnar/vaultvoice.git
cd vaultvoice
npm install
npm run buildOption A: Ubuntu/Debian (without virtual environment)
pip3 install --break-system-packages librosa numpy scipy soundfile pydub
pip3 install --break-system-packages --no-deps resemblyzer
pip3 install --break-system-packages torch --index-url https://download.pytorch.org/whl/cpuOption B: Any system (with virtual environment — recommended)
python3 -m venv .venv
source .venv/bin/activate
pip install librosa numpy scipy soundfile pydub
pip install --no-deps resemblyzer
pip install torch --index-url https://download.pytorch.org/whl/cpuWhy 3 separate commands? Installing everything at once downloads ~3GB of GPU libraries you don't need. This way it's ~200MB total.
Warnings about
webrtcvadandtyping? Ignore them. VaultVoice doesn't need them.
That's it. You're ready.
Linux:
arecord -d 10 -f S16_LE -r 16000 my-voice.wavmacOS:
sox -d -r 16000 -c 1 my-voice.wav trim 0 10Windows: Use any voice recorder app, save as .wav.
node dist/cli.js register --name "my-voice" --audio my-voice.wavnode dist/cli.js verify --profile "my-voice" --audio suspicious-call.wavOutput:
🔍 Analyzing: suspicious-call.wav
Speaker Match: 100.0% (threshold: 85%)
Deepfake Score: 0.58 (medium risk)
Verdict: ✅ AUTHENTIC — Likely real voice matching profile "my-voice"
Details:
Duration: 10s
Quality: Good (SNR: 40dB)
Spectral: Natural harmonics detected
Artifacts: None found
node dist/cli.js scan --audio unknown-audio.mp3node dist/cli.js list # Show registered profiles
node dist/cli.js delete --name X # Delete a profileSpeaker Verification: Extracts a 256-dimensional voiceprint using resemblyzer (neural speaker encoder). Compares via cosine similarity — above 85% = match.
Deepfake Detection: Analyzes 5 spectral features that AI-generated audio gets wrong:
| Feature | What it catches |
|---|---|
| Spectral flatness | Synthetic audio is too uniform across frames |
| Pitch consistency | Cloned voices lack natural pitch variation |
| MFCC patterns | Higher-order coefficients reveal synthesis artifacts |
| Zero-crossing rate | Natural speech has characteristic variation |
| Spectral rolloff | Vocoders produce distinctive patterns |
Score above 0.65 = likely deepfake. Everything below = clean.
.wav .mp3 .ogg .flac .m4a
Non-WAV files are automatically converted to 16kHz mono WAV. Requires ffmpeg.
- All processing is 100% local
- Voiceprints stored in
~/.vaultvoice/voiceprints.db(SQLite) - Zero network requests — nothing is ever sent anywhere
- No telemetry, no analytics, no tracking
| VaultVoice | Pindrop / Nuance | |
|---|---|---|
| Price | Free | $$$$$ |
| Privacy | Fully local | Cloud-based |
| Setup | 5 minutes | Weeks |
| Target | Individuals | Call centers |
| Accuracy | Heuristic (good for personal use) | ML-based (higher accuracy) |
- 🔒 Personal security — verify voice messages from family are real
- 🎙️ Content creators — detect unauthorized clones of your voice
- 📰 Journalists — authenticate audio evidence
- 💼 HR — verify identity in remote interviews
- ⚖️ Legal — screen audio evidence for manipulation
# Remove project
rm -rf vaultvoice
# Remove stored voiceprints
rm -rf ~/.vaultvoice
# Remove Python dependencies
pip3 uninstall resemblyzer librosa numpy scipy soundfile pydub torch -ynpm test # Run tests (30 tests)
npm run build # Build for productionMIT