First problem: Your system needs to be perfectly synced up in time with the known sound being played. Human hearing goes to ~20KHz, so you need to be synced up with an error of <0.1 ms per second, probably much less. That’s gonna be tough.
Next problem: as you move around the room, your distance to the speaker changes, and you go out of phase. So you need positional data for your ears with sub-centimeter accuracy.
Can we get rid of the first two problems by using a microphone to detect the signal and auto-correct its phase as you move? Maybe but then you need an incredible algorithm to be able to pick out the correct part of the signal from all the background noise and echoes, plus it has to run in real-time at high sample rates. Very very difficult.
Then even if you solve these issues, you also have to cancel the echoes of the song bouncing off the walls etc. If you’re trying to take advantage of the fact that it’s a “known song” you then need an extremely good model of the room you’re in to predict how the sound will bounce.
In short, I suspect it’s “theoretically possible” but completely infeasible.