At Herd Security, I led the development of an AI-powered voice authentication system to help businesses detect if a caller’s voice was real or AI-generated-preventing scams like impersonation fraud. Here’s how we did it in simple terms:

1. Understanding the Problem

• Scammers were using AI voice clones to trick call centers (e.g., pretending to be a CEO or customer).
• We needed a way to instantly flag fake voices without slowing down calls.

2. Gathering Voice Data

• Collected real human voices (from public datasets and anonymized call recordings).
• Generated fake AI voices (using tools like ElevenLabs and Resemble.AI) to train our system.

3. Building the Detection System

• Step 1: Analyzed voice fingerprints (e.g., tone, pacing, background noise).
• Step 2: Used AI models trained on real vs. fake voices to spot differences.
• Step 3: Made it fast-under half a second per check-so call centers could use it live.

4. Testing & Deployment

• Tested with banks and customer service teams to ensure accuracy.
• Integrated into call center software (like Zendesk) and mobile apps as a lightweight security feature.

5. Results

• Detected 91% of high-quality AI fakes (like cloned CEO voices).
• Continuously improved by learning from new scam tactics.

Why It Worked

• Speed: Instant checks without disrupting calls.
• Accuracy: Few false alarms, so businesses trusted it.
• Adaptability: Evolved as scammers improved their tech.