Voice Guard

Inspiration

Imagine getting an urgent call from a loved one saying they need money fast. The voice sounds 100% real, because scammers can now clone voices in seconds using AI tools. AI voice impersonation and push payment fraud are exploding into a multi-billion-dollar threat. The real issue? Traditional banking apps check if a user logged in and tapped "send," but they can't tell if the user was manipulated into doing it. They only verify the sender, leaving the recipient completely unverified. We designed VoiceGuard to fix this broken trust loop before money leaves an account.

What it does

VoiceGuard is a dual-sided fraud prevention framework for mobile banking. When a high-risk payment starts, VoiceGuard holds the funds and launches separate in-app calls to both the sender and recipient:

1: 3-Word Challenge: Both parties repeat a fast-expiring prompt (e.g., "maple, river, lantern"). 2: Audio Watermarking: DeepMark technology analyzes the live audio stream for synthetic speech or manipulation. 3: Mutual Approval: Both parties review transaction details in-app and confirm. If either fails or declines, funds stay on hold.

How we built it

1: In-App Digital Calls: Routing calls inside the banking app to bypass unauthenticated cellular networks. 2: Ephemeral Prompts: Generating short-lived, transaction-tied 3-word challenges. 3: DeepMark Pipeline: Embedding DeepMark's SDK directly into the streaming backend for real-time audio verification. 4: Payment Gate Engine: Integrating with core banking APIs to hold, release, or flag pending transfers.

Challenges we ran into

1: Audio Latency: Keeping real-time speech verification and watermark checks under low latency. 2: Edge Cases: Cleanly handling dropped calls, timeouts, and uncooperative recipients without locking user accounts. 3: Low Friction: Making the 3-word prompt feel fast, natural, and reassuring.

Accomplishments that we're proud of

1: Solid Architecture: Designed an end-to-end security framework that verifies user intent, not just passwords. 2: Smart Security Integration: Paired real-time DeepMark audio watermarking with dynamic speech verification in a standard banking workflow. 3: Bank-Ready Vision: Created a go-to-market strategy aligned with growing regulatory pressure to stop push payment fraud.

What we learned

1: Authentication != Intent: Single-device logins can't stop users from being coerced by AI voice clones. 2: Layered Defense Wins: App sessions, dynamic challenges, and audio watermarking create an exponentially stronger barrier together. 3: High Market Demand: Financial institutions urgently need targeted friction tools to curb non-recoverable fraud losses.

What's next for Voice Guard

1: Build Working Prototype: Develop the backend and UI for live payment gating and dual in-app verification calls. 2: DeepMark SDK Integration: Embed DeepMark’s SDK into our streaming backend for real-time testing. 3: Banking SDK & Pilots: Package VoiceGuard as a plug-and-play SDK and target regional credit unions for pilot trials.