Aside from voice quality (Tortoise is undefeated) and speed (Tortoise is vastly slower), there is a difference in consistency and control.

Tortoise is a probabilistic system that uses one model to randomly generate 80+ different possible outputs and then another model to pick the best one. A consequence of this approach is that sometimes none of the generated outputs are good, the selected 'best' choice is weird and you have to run it again to get something you like.

(That's not necessarily a bad thing. It's actually incredibly cool to be able to say "that's not quite what I want, can you say the same thing but differently?" and have the computer just keep saying it in different ways until you're satisfied. But, it's probably not the right choice for something like Siri where you want consistently decent output 100% of the time, without human review.)

Tortoise also currently lacks any direct control over pronunciation or speaking rate, which is a deal-breaker for many applications.

This not only does not require a GPU, but also works on 1-4 CPU threads (!):

- 8 kHz, 1 thread 15-25, 4 threads 30 - 60 - 24 kHz, 1 thread 10, 4 threads 15 - 20 - 48 kHz, 1 thread 5, 4 threads 10

the numbers are seconds of audio generated per second