Update to work with mps, tested on M5 Max

[taxilian]

I don't know if this is the best way to do all of this, but it's working for me! Fixes #1

[taxilian]

Update: while this does work, it doesn't seem to keep a consistent voice between sentences and I can't figure out why. I think there is likely a bug somewhere further in than the code in this repository, but I don't understand the infrastructure well enough to be sure.

If I use cpu inference it works, but using mps a cloned voice works only for the first sentence and then changes voices (entirely changes the voice, not just some drift) to an apparently random voice each sentence thereafter.

[AodenTeo]

Hey @taxilian, thanks so much for helping with this. I'll have a look and try to see what the issue is.

[taxilian]

With that fix I'm not having the issue with it changing voices anymore; it does take it to a float32, though, not sure what impact that has on anything.