Just create a series of files each with it's own audio for a single letter, then string them all together and send the concatenated audio.

But wouldn't it then be easy to generate a program that knows about the signature of each of those letters? The challenge is randomizing the sound of each letter so that it is easy to understand for a human, but hard to recognize by a program.

This is especially tricky since there is already some really good speech recognition software out there.

Perhaps you can overlay the sound of the letters/words with some music, or something else that will not affect the ability of the person to hear the word.