I don't have any experience with this, but I'm a physicist with some linguistics background, so I can make some guesses. Phonologists analyze speech by looking at the formants [1]: spectral components of the sound that are (typically) higher in frequency than the fundamental (i.e. pitch). The formants are determined by the speaker's mouth and tongue shape, while the fundamental is set by the vocal chords. If you want to change the pitch of a speaker's (or singer's) voice but not change the speed or the phonemes (fundamental speech sounds), then you need to change the frequency of the fundamental while preserving the spectral envelope (on the harmonics of the fundamental) set by the formants.
How to actually implement this becomes a question of signal processing. I have an idea of how I might do it, but it would be pretty rough. I tried some other Google search keywords and found a PDF [2] that gives one approach and might be a good place to start.
[1] http://en.wikipedia.org/wiki/Formant
[2] http://dsp-book.narod.ru/Pitch_shifting.pdf