Anti-Deepfake On-Device
A voice-liveness classifier runs on-device — on the phone today, moving to the earbud's dedicated NPU as that hardware ships — never in the cloud. It shows a real-time confidence score during every call, including unencrypted ones (Teams, Zoom, WhatsApp, traditional telephony).
Why on-device, not cloud
Sending voice to a cloud service for authentication introduces three problems: (1) you transmit the very signal you want to protect, (2) the provider can analyze it at its own discretion, (3) latency incompatible with natural conversation. Q-AUDION solves all three by running inference locally — on the phone today, on the earbud's NPU as that hardware ships.
- No audio ever transmitted
- Runs alongside the call in real time
- Zero dependency on external services
Built on a peer-reviewed anti-spoofing architecture
Guardian Mode uses AASIST, the reference architecture for the ASVspoof anti-spoofing benchmark — trained to recognize speech-synthesis, voice-conversion, and replay attacks. It runs quantized for real-time, on-device inference.
- Public, peer-reviewed architecture
- INT8 quantization per deployment embedded NPU
- Deployed today on the phone; migrating to the earbud's NPU as hardware ships
Confidence score, shown in real time
During every call, a confidence score (how likely the remote voice is synthetic) is shown in the call UI, with a warning when it drops below a safe threshold. It works on any call, encrypted or not.
- Recalculated continuously, frame by frame
- Runs for the full duration of the call
- Threshold configurable via Sovereign Server policy