The Limits of Deep Learning Models

One of the things that we were really worried about is that these deep learning models, it's really hard to disentangle features of what they've learned. Your accent has a lot to do with your emotion. So de-accenting a particular speaker or replacing an accent of a particular speaker kind of replaces quite a bit of the style of that speaker as well. You could prompt the model in interesting ways such that you could retrieve a voice that is proprietary. Just for example, we crawled all over YouTube, gathered data sets from celebrities and etc. And somewhat at their best impersonation of Morgan Freeman, the model would most likely just spit out Morgan Freeman because that's like the closest