It's Kafkaesque source material not withstanding, it was able to accurately emulate two of the most robotic sounding popular 'influencers,' with tons of ocntent to piece it all together, but and this is a caveat, the only realistic thing that came from it was when they briefly spoke over each other. Everything else seemed obviously fake.
Consider this sounds nothing like the Elon you see on the most recent All in Podcast interview, either. If you threw that in would it distort the pitch and cadence of 'fruit juice' Elon?
Over all it's a pretty cool display of where the tech is, though, but NLP has had decades of data behind it but it's still in its relative infancy. The sounds of the voices being made within a situational context are passable deepfakes, for sure. It's almost like they got audio of both together to make a SNL sketch with a deadpan hook that never got aired.
Conclusion: Model training is actually really cool and has thus far been the most interesting aspect of AI to me.
It's prompted by a premise of the podcast that I wrote and a list of keywords and guests. I also sometimes prod it during the transcript generation to make it more interesting, the ending for example was me injecting a line "I'll leave you with this bit of Vogon poetry".
https://soundcloud.com/jn2022/lexman-artificial-podcast-epis...