The "Void Parity" is the issue that occurs when solving anything equivalent to a 3x3x3, particularly a void cube, where you have odd total parity with your current assignment of centers.
Layer methods hit this (i.e. you can recognize the parity) at PLL, and I think with CF or Roux, you can try to avoid it.
It's well-known that a reassigning centers with a quarter-slice twist + resolving pieces will resolve this. My approach used to be something like an M-fix with a 4-flip and two 3-cycles.
But for speedsolving, you'd want something efficient. Tomas once asked us for a fix in #rubik, and so I got kinda interested.
The most practical is to find a fix that preserves OLL (which can be applied before PLL). It's not hard to imagine that the best fixes are URrML-ish and preserve CPLL, which is also good for CF and Roux, depending on how you do it (and what kind of mistakes you make).
After a while, I found l U r' U M U' M' r U' R' U M U'.
Then, qqwref found the very nice M' U M' U M' U' M' U' M' U2' M' U' M'.
So, does anyone have any other algs, or better approaches?