Why Long Songs Are Difficult

A model can generate ten convincing seconds without knowing how to organise three convincing minutes.

Short clips mainly test whether the sound is locally believable. Do the instruments fit together? Does the rhythm continue naturally? Does the next chord make sense?

A complete song creates additional demands. A chorus should return recognisably. A second verse should relate to the first without copying it exactly. Transitions should lead somewhere. Energy should rise and fall. The ending should feel connected to what came before.

Long-form generation is therefore not only short-form generation repeated for longer.

Duration is not structure

A system that produces 180 seconds of audio has solved a duration problem. It has not necessarily solved song organisation.

Length answers how much audio exists. Structure answers questions such as:

1. Which sections exist?
2. When do they begin and end?
3. Which material returns?
4. What changes during a return?
5. How does energy develop?
6. How does the song conclude?

Stable Audio research showed that training on long temporal contexts can support full-length outputs. Whole-song symbolic research goes further by representing form, phrases, cadences, chords, and notes at separate levels.

The context window becomes musical memory

As Section 2 explained, a model predicts from the information available in its context window.

For long music, the important question is not only how many tokens fit. It is how many seconds, bars, or sections those tokens represent.

A context of 4,096 tokens may cover a long symbolic score if each token represents a note event. The same number may cover much less time when several codec tokens are required for every fraction of a second.

Longer context costs more

Standard full self-attention compares many sequence positions with one another. Its memory and computation grow quickly as the sequence becomes longer.

Music Transformer introduced an efficient relative-attention formulation to improve long symbolic generation. Museformer later combined fine-grained attention to selected structure-related bars with coarse summaries of other bars.

These approaches do not merely increase the context. They change which earlier information receives detailed attention.

Music needs repetition, but not endless copying

Repetition helps listeners recognise motifs, grooves, sections, and identity. A chorus should sound related to its earlier appearance.

The difficulty is balancing three behaviours:

1. Return: recognisable material comes back.
2. Variation: something changes in orchestration, melody, rhythm, harmony, or energy.
3. Development: the return contributes to a larger direction.

A model that avoids repetition may sound aimless. A model that repeats too exactly may sound trapped in a loop.

Musical identity must survive change

The identity of a song can come from several sources:

1. A melodic motif
2. A chord progression
3. A drum groove
4. A vocal phrase
5. A particular instrument combination
6. A production texture
7. A repeated section pattern

Long-form generation must preserve enough of these features to keep the same piece recognisable while allowing the arrangement to evolve.

Prompt adherence can weaken over time

A text condition may describe the overall style, but the generated sequence also becomes a strong source of context.

As the output grows, the model may increasingly follow patterns in its own recent audio. Instrumentation, energy, tempo, or mood can gradually move away from the original request.

This is related to the path dependence described in Section 2. Long-form generation gives those small deviations more time to accumulate.

Transitions connect local quality with global form

A verse and chorus can both sound convincing in isolation while the movement between them feels abrupt.

A useful transition may prepare several features at once:

1. Harmonic movement
2. Rhythmic density
3. Instrument entrances and exits
4. Vocal phrasing
5. Energy and loudness
6. Effects and production texture

The transition must reflect both where the song has been and where it is going.

Generating in chunks

One way to create long audio is to divide it into shorter regions. Each new chunk can overlap with, continue from, or receive a summary of the earlier material.

Chunking reduces the amount generated at once, but introduces new problems:

1. Audible seams between chunks
2. Repeated material near overlaps
3. Changes in timbre or loudness
4. Lost motifs from distant chunks
5. Weak awareness of the complete song plan

Chunking solves a memory or computation problem, not automatically a structure problem.

Endings require awareness of remaining time

A locally trained model can always produce another plausible bar. That does not tell it when to stop.

A convincing ending may require:

1. Completing a phrase
2. Resolving harmonic tension
3. Returning to a tonal centre
4. Reducing or concentrating energy
5. Repeating a final hook
6. Applying a cadence, cutoff, fade, or outro

Duration conditioning tells a model how much time remains, but the training data and architecture must teach what endings sound like.

Explicit structure before melody

MELONS represents melody structure as a graph containing several types of bar-level relationships. It separates the task into structure generation and structure-conditioned melody generation.

This approach makes repetition and transformation relationships part of the representation instead of expecting them to emerge only from a long token history.

Planning from form to notes

Whole-Song Hierarchical Generation models symbolic pop music through several levels. Higher levels describe phrase structure, key progression, form, and cadences. Lower levels represent chords, accompaniment, and notes.

The levels are generated from the top down. Each lower level receives the higher-level plan as a condition.

This makes whole-song organisation an explicit generation target rather than a side effect of local note prediction.

Evaluating a complete song

Long-form evaluation should examine several levels separately.

A useful review can ask:

1. Is the audio locally clean?
2. Are tempo, key, instrumentation, and production stable where intended?
3. Do motifs and sections return recognisably?
4. Does repetition include useful variation?
5. Are transitions prepared and resolved?
6. Does the energy trajectory change meaningfully?
7. Does the output follow its prompt throughout?
8. Does the ending feel deliberate?

One score rarely captures all of these properties.

From architecture to training

This chapter explained how music can be generated through prediction, autoregression, hierarchy, diffusion, and conditioning. It also showed where these systems fail over long durations.

The next chapter, Training and Fine-Tuning a Music Model, examines how model parameters learn these behaviours. It covers objectives, optimisation, fine-tuning, data mixtures, and the trade-offs involved in teaching a model to follow prompts, preserve structure, and generate convincing sound.