I suggest using 8-bit sampling at 8kHz sampling rate. Use a 4kHz low pass filter to meet the Nyquist criteria. That would approximate telephone quality voice(not as good because of no CODEC). That would require 8k of memory for each second of voice. Ex. 60 seconds = 480k of memory. You could try reducing sampling rate and filter cutoff frequency until the voice quality is ~JUST~ what you need, that would save on memory requirements. Make sure you have a stable timebase for sampling, or the voice quality will suffer. I haven't tested this, but this technique is widely used. Sounds like fun. I would still probably use an ISD device, though they are pricey. Use standard parallel interface ADC and DAC.

Cory Spackman