Depends on how the voice works. If its a phoneme system that can say anything at all, then its always going to sound robotic. If its a real person and it only has whole words digitised for specific places, like the Wallace an Grommet clocks for example, then you would need only to make a chip with the right coding for the device with other voices on it. Certainly the old Pure Sonus 1 DAB had that latter method and also allowed two choices by downloading voices from the website, sadly they never made much of it, just a male and a female were available.
So really you need to find out how its done and would need to know how the indexing is done and word assembly if its phoneme based,and record your own phonemes for every sound needed for speech which is very long winded, and indexing of reserved words if its the whole word kind. Brian