Amin

Deepfakes defeat ears. They don't defeat hardware. Voice cloning now needs only seconds of audio, and it is being used for "family emergency" and fake bank-officer calls that walk victims into UPI payments while still on the phone. Statistical deepfake detectors are the wrong defense: probabilistic, always one model behind, and they produce no proof a user can act on mid-call. Wavetone proves who is really on the call — cryptographically: 1) OUT-OF-BAND (hardware proof): ECDSA P-256 challenge-response signed inside Android StrongBox/TEE, relayed over a lightweight WebSocket. Verdict: HARDWARE VERIFIED. 2) IN-BAND ACOUSTIC (paired-device proof, zero internet): a pure-Kotlin BFSK modem (1600/2400 Hz, ~640 ms burst) carries a truncated HMAC-SHA256 + CRC-8, decoded by a Goertzel engine — no native DSP libraries. Verdict: PAIRED DEVICE VERIFIED, labeled honestly (paired-secret possession, not enclave proof). 3) SCAM-BREAKER: a CallScreeningService seeds sessions for real calls. If proof fails or the caller is unpaired, an AccessibilityService locks UPI/banking apps ("Pay ₹", "Enter UPI PIN") behind a crimson shield until the call ends (auto), the user ends it from the shield, or holds a 3-second emergency override. Sessions are AES-256-GCM encrypted and persisted, so protection survives process death mid-call and lifts itself on hang-up. Built and verified end-to-end: hand-rolled SHA-256/HMAC and DSP validated against NIST/RFC vectors; 35 unit tests; 6 instrumented tests on a physical Pixel 9 Pro XL (StrongBox signing, true over-the-air speaker→mic decode, replay rejection, ciphertext at rest); Lint at 0 errors; 2.4 MB release APK; CI on every push.