Generated music has become hard to fault across a few bars and remains easy to fault across a full track. The gap is between local plausibility and long-range structure, and it is the central unsolved problem in the field.
Prediction is local by construction
These systems generate audio by predicting what follows from what has just occurred.
That objective produces excellent continuity at the scale of a phrase, because the immediately preceding material contains most of the information needed.
Song form operates at a scale of minutes, and nothing in the local objective requires a chorus at two minutes to relate to the one at forty seconds.
Repetition in music is exact, not approximate
A returning chorus in a conventional song is usually the same material, recognisably identical rather than merely similar.
A model generating forward produces something in the same character, which is a variation rather than a return.
Listeners register the difference immediately, because recognising a repeat is how they track where they are in a piece.
The sections stop earning their place
Structure works by contrast: a verse establishes something, a chorus lifts, a bridge withholds and delays the return.
Those relationships require a plan held across the whole track, and a system generating sequentially has no representation of a plan.
What emerges instead is continuous variation at roughly constant intensity, which is why generated tracks often feel like they could end anywhere.
Conditioning on a structure helps and constrains
The working response is to supply the form externally, specifying section boundaries and lengths and generating within them.
This produces recognisable songs and moves the compositional decision back to the person, which is a reasonable division of labour rather than a defeat.
It also means the system is arranging rather than composing, and the difference matters for how the output should be described.
Long context is expensive in audio
Audio carries far more data per second than text, so a model attending across a full track handles a very long sequence.
Approaches that compress audio into coarser representations before modelling make the span tractable and discard detail that matters to timbre.
The trade between how much of a track a model can consider at once and how well it renders any given second of it is the constraint that current architectures have not escaped.