How pitch over time is measured
The detector cuts the sound into short overlapping frames, about 60 times a second for the microphone and every 16 milliseconds for a file. For each frame it uses the YIN method, an autocorrelation-style technique that looks for the shortest repeating period in the waveform and converts it to a frequency in Hz. The same approach powers the tuner and vocal range test on this site. A confidence value tells how clearly periodic the frame was.
Frames that are too quiet or not clearly periodic are treated as unvoiced and left out of the plot. The remaining pitches are converted to note numbers relative to your chosen A4, then smoothed with a short median filter so single wrong frames, such as an octave jump, do not draw spikes. The result is a pitch contour drawn on a piano-roll grid with note names on the vertical axis and time across.