Pitch detection answers what note, and it is harder than it sounds. Real instruments stack harmonics that can outweigh the fundamental, mixes contain many notes at once, and octave errors lurk everywhere. Monophonic detection (one voice) is solved enough to power tuners and Auto-Tune; polyphonic detection remains genuinely difficult.
Classic methods include autocorrelation and the YIN algorithm in the time domain, and harmonic pattern matching on spectra. Modern systems increasingly use trained models.
Reactive visuals mostly sidestep the problem: bands and onsets carry the show, with chroma as the practical middle ground when harmony should matter. Full transcription is rarely worth its latency on stage.