VoiceGuard

Problem

Number-based spam filters have a blind spot: they can’t detect a caller whose voice is synthesised on the fly. Modern voice-phishing attacks use text-to-speech and voice-conversion tools that no blocklist will ever catch. The threat arrives looking like a human and sounding like one.

VoiceGuard is a research engine that moves the detection down to the audio stream itself — running entirely on the device, with no data leaving it.

Approach

The engine combines three complementary signals to distinguish a human caller from an AI one:

• Behavioural latency — real conversations have irregular turn-taking gaps; AI pipelines tend to be suspiciously consistent.
• Background-noise linearity — synthesised audio often has artificially clean silence or a looped ambient track.
• Spectral artifacts — a lightweight TensorFlow Lite model flags the anomalies that neural vocoders leave in the frequency domain.

None of these signals is reliable alone. Together, with a short analysis window, they push detection accuracy above 85% at under 50 ms per chunk.

Architecture

The engine follows a hexagonal architecture in Kotlin, with Coroutines and StateFlow keeping the audio pipeline non-blocking. The ML model runs on-device via TensorFlow Lite — no network call, no latency budget eaten by a round trip.

Status

Currently specification-first: the product requirements document is guiding the initial R&D phase, with the detection pipeline under active development.