AI music systems produce a range of musical output. At the smallest scale, a system can generate a melody: a single line of pitches with some rhythm. It can generate accompaniment, such as chords or a bass line that supports a melody someone else wrote. It can generate a full song with several instruments playing together. And it can produce audio, the actual sound you hear through speakers.
A useful split runs through all of these. Some output is symbolic: it is written as musical instructions, like which note to play, when, and for how long. A melody stored as note names and durations, or a chord progression stored as chord labels, is symbolic. Other output is audio: a waveform that directly represents pressure changes over time. Symbolic material is compact and easy to edit, because you can change a single note. Audio is what you actually listen to, but it is much harder to edit note by note.
The same piece can exist in both forms. A symbolic melody becomes audio once it is played or rendered by an instrument or synthesizer. That conversion step matters, because a system that works in notes and a system that works in sound are solving different problems.