It can run on CPU without much issue and takes up a few gigs of RAM and will produce about in realtime. If you GPU accelerate you only need about 8GB of video memory and it will be at least 5X faster.

Out of the box it's not as good as Eleven Labs based on their demos, but those are likely cherry picked. There are some tunable parameters for the Bark model and most consider the output high enough quality to pass into something else that can do denoising.