Transcription sounds like magic until you see the pipeline. Audio goes in as pressure measurements tens of thousands of times per second; notes come out as discrete events. Between them sit four stages, each making judgment calls.
Windows and spectrograms
First the audio is cut into short overlapping windows, and each window is transformed into a spectrum: which frequencies carry energy at that instant. This is where time resolution trades against pitch resolution — long windows hear pitch precisely but blur timing.
Onsets, frames, contours
A neural model then reads the spectrogram three ways: onset frames (a note begins here), note frames (a note sustains here), and pitch contours (the fine frequency path between). Combining the three is what separates a transcription from a piano-roll-shaped blur.
The judgment calls
Finally thresholds decide everything you hear: how much onset evidence justifies a note, how short a note may be before it's noise, whether an octave harmonic becomes its own note. When a transcription disappoints you, it is almost always one of these dials — which is why a good converter lets you turn them yourself.