Inspired by the surge in deep-fake phone scams, we built a five-gate voice-authentication workflow that silently verifies a caller in <5 s: Twilio + VAPI capture a secret passphrase, a FastAPI microservice computes ECAPA embeddings and scores them with $\text{score}=\frac{v_{\text{live}}!\cdot v_{\text{enroll}}}{\lVert v_{\text{live}}\rVert\lVert v_{\text{enroll}}\rVert}$, raw-waveform CNNs flag deepfakes/replays, Duo push adds human confirmation, and Supabase row-level tokens release data only when all gates align, all orchestrated in n8n. We learned to balance user experience with security—hitting $\text{FAR}=\text{FRR}\approx0.01$ demanded int-8 quantization, WebSocket back-pressure, and augmented ASVspoof datasets—while overcoming GPU-quota limits, latency spikes, and sparse anti-spoofing corpora. The result: a plug-and-play, defense-in-depth template that turns any phone call into a castle with five doors that snap shut the moment you say “hello.”
Authy
Our hackathon-ready workflow captures a passphrase, verifies your unique vocal fingerprint with SpeechBrain, sniffs out deepfakes and replays, and only then token-opens your data
Updates
Leave feedback in the comments!
Log in or sign up for Devpost to join the conversation.