Here is another discussion by Dan York on the VoIPSA Blog and Bluebox Podcast about using covert channels to tunnel data in a voice call. The research describes using 8kb (of the 64Kb) in a voice call to carry some sort of covert data or communications. This of course would require an application working at both ends to code and decode the signal, which would probably need to be encrypted or obfuscated in some way to avoid detection. I scanned the actual research paper and it looks like they are mainly using low order bits for the channel, which presumably wouldn't affect the audio. Perhaps you could also send packets with out of order sequence numbers, which if done right, would be rejected by target jitter buffers, but could be used to carry data.
Of course the real problem here is bandwidth - it would take a long time and a lot of calls to carry your average word or powerpoint document.
Here are the direct links: