The Si3000 has been pretty good. It is a high quality 16-bit mono voice codec in a reasonable pkg to solder. It has on-board Programmable Gain Amplifiers and lots of neat stuff. You do not need to do any input or output filtering to remove components above the normal 4KHz voice range.
Its output format is a bit difficult to use unless you have a DCI or Framed SPI port. The dsPIC33/30 or PIC24 have them, and those are also 16-bit cores so they handle the data words well. The PIC18F series, not sure if the SPI could be configured to read in, in which case you're out of luck.
The dsPIC30-33 series also have a software "Speex" encoder/decoder to compress voice down to 1KB/sec. It can compress and decompress on the fly.
I have a ready-to-go driver project to interface 1 to 3 Si3000 devices with a 33F using the DMA controller. DMA controller is nifty, allows the module to fill a buffer without interrupting the core and can even operate when the core is in SLEEP.