To be clear, it's my understanding that playing music or generating artificial noise won't actually do much to stop this attack. Listening to some good tunes will definitely help distract you from worrying about attackers stealing your private keys, though!
It depends on which attack vector you're defending against.
If you're trying to defend against the cellphone (or chasis microphone) attack vector, then you can defintely defend with human audible noise. Your cellphone is going to have an aggressive low pass filter prior to the ADC to prevent aliasing. This will limit the cellphone range to roughly human audible range. The difficulty is in generating sufficient noise at the frequencies required. As the FAQ on the page points out the majority of this lies in +10kHz, so using normal music won't be too effective, but artificial noise is certainly possible.
The Q&A on the linked site answers all sorts relevant questions and is quite enlightening.
In the actual paper is maddening non-specific about how widespread audible range attacks are. Ultrasonic attacks are likely viable against nearly all unprotected machines, however the paper only says that 'some' machines are vulnerable to audible spectrum attacks.
If I understand correctly, it's because the frequency it is listening to is much higher than normal music. Also possibly because the algorithm is meant to deal with noise.
Actually it is because it is all fake, so he can set whatever rules he wants.
If you are listening for sounds higher than "normal music" you can't capture them with a standard mic, let alone the mic on your phone, since it is optimized for vocal range.
