How Deep Neural Networks Can Detect Deep Fake Speech

We've trained a generalized deep fake detector where it only works on audio. It's about 80% accuracy right now with blind tests. And then as you mentioned with these models, all the fine tuning work that we've done in the past, since we're so good at it, you can get about 98% accuracy if you just show the model those speakers. So if you wanted to protect your organization, we would collect data from those 100 speakers and then fine tune that model. That's also particularly exciting.