A project I worked on in college was a "steganographic modem", transferring data in the audio spectrum by lacing a song being played with a signal, such that the signal in each frequency band would always be below the threshold that the human ear could tell the difference.

Even in a controlled setting the bitrate we achieved was very low, but for most songs it was enough to encode the lyrics. Seems quite useful until you remember both ends already need the original audio file in its entirety!