Encoding it would be easy. I'd use a total internal reflection switch like this:
http://www.google.com/patents/US6056696
The voice would vibrate the membrane, which would modulate the light by passing it, or not passing it as appropriate, much like an electrical microphone. The hard part would be manufacturing it to the required tolerances, which means such phones would be rare, and expensive, limiting their utilization. The Morse code versions would be much more common, then.
What would be much harder is how you'd use the light to reproduce sound on the other end.....With electricity, you could record it with a film strip, and play it back, like they did with the early talking films.
http://en.wikipedia.org/wiki/Sound-on-film
Could you maybe record the output on a vinyl disc, and play it back with a mechanical phonograph? You might need some kind of photoresist/acid etch system to create the grooves. This makes sending messages easy, but receiving and decoding them hard.
http://en.wikipedia.org/wiki/Photoresist