I think that 16 kHz / 16 bit recordings should be sufficient for the development of a speech model. But what does that mean? A good article explains the differences between amplitude (16 bit recordings are more precise than 8 bit recordings), and frequency, wavelength (the human ear can distinguish up to 20 kHz; you need the double amount of kHz for recording; 16 kHz means that 8 kHz are distinguished - should be sufficient for speech).