If you're familiar with the 56k modem connection song, that's a similar principle. The telephone line is analog but digital content can be encoded in analog using a DAC, at which point it's can be interpreted as an audio signal and played through a speaker.
The sound itself was just data (training patterns, not the actual data of interest, that gets transmitted after the "song") being encoded at a faster and faster rate, and it would select the fastest speed it could transmit data at reliably along the line. The "song" would progress further the higher the speed it could communicate reliably with the other side, or it might go back and repeat a section because it could negotiate phase (and other parameters) as well as bitrate, which might improve the line reliability, but it still needed to see that test pattern come through cleanly before it would accept a rate.