AI voice-clone scam detector
Real or Clone?
Problem
Scammers can clone a voice from a few seconds of audio and send an urgent voice note asking for money, and most people cannot tell the clone from the real person.
Objective
Give anyone a fast, honest second opinion on a voice note, and measure how well it holds up on voices and generators it has never seen.
Solution
XLS-R 300M fine-tuned on public real and fake speech plus 1,619 fresh Chatterbox clones generated on an NVIDIA L40S, with real and fake audio both put through the same WhatsApp-style compression. It is served by FastAPI behind a mobile-first app. Held-out testing caught a shortcut in the first model, which flagged most real voices from outside its training data. A third run with more varied real voices fixed it: 96.1% accuracy, 2.4% EER on unseen generators, and about 7% of real voices wrongly flagged.
Main features
What the product includes.
- Real / clone / uncertain verdict with confidence
- Timeline of suspicious seconds
- Upload or record a voice note
- LLM safety tips written from the verdict only
- Before/after fine-tuning model switch
- Private per-user history and dashboard
- Arabic, French, and English with RTL
Technology
Built with the right tools for the job.
Challenges
Balancing the project’s technical requirements with a clear, responsive experience—and keeping the implementation maintainable as the scope grew.
What I learned
The project strengthened my ability to turn a broad idea into structured features, make deliberate technology choices, and communicate complex functionality through a polished interface.
Next project
Have a product that needs thoughtful execution?
Tell me what you are trying to build, who it is for, and what success should look like.
Start a conversation