Editing Music Without Pretending to Edit Waveforms

YuE2 makes a musical edit meaningful by giving it a symbolic home before generation begins. This chapter turns that idea into a bounded method: preserve an untouched baseline, state which musical facts must remain fixed, change only the intended score or request fields, and regenerate a complete recording. The goal is not to promise a surgical waveform patch. It is to make a composition change specific enough to inspect, test, and describe honestly.

The difference between “change the harmony” and “make it sound more jazzy” is the difference between a contract and a mood. The first can identify what should vary and what should not. The second leaves melody, timing, form, instrumentation, and even the meaning of success unresolved. YuE2's score interface cannot remove the need for musical judgment, but it can put that judgment into a representation where a person or agent can work with it.

Start with a contract, not a vague edit request

A bounded edit begins by freezing the baseline. Generate or import the original composition, retain its score and audio, and do not overwrite the directory that will serve as the comparison point. The saved plan contains the human-readable score when applicable, the exact encoded ABC IDs and prefix, and an integrity manifest. Loading it verifies that the score and arrays still agree. This protects the distinction between reusing an unchanged plan and introducing an intentional intervention.

The next step is to state the invariants. An invariant is a musical fact that the edit is not allowed to change. Depending on the goal, the fixed set might include the vocal melody's pitches and timing, the instrumental theme, lyric order, meter, tempo, section order, or only the broad melodic contour. The variable set might include chord symbols, style, instruments, lyrics, tempo, or form. The contract should name the relevant voices and passages rather than assuming that “the song” is a sufficiently precise object.

This level of specificity is not bureaucratic overhead. A full YuE2 plan can expose vocal and instrumental melodic material, timing, structure, and harmony. If a user asks for a reharmonization, the useful claim is not that the system “edited the song.” It is that a new symbolic plan changed the intended chord material while preserving the declared melody and structural invariants. If the user also changes lyrics or tempo, that is a different contract and may require different checks.

A copied score is the safest carrier of the intervention. The documented workflow is to copy the ABC, edit the copy, inspect it, and supply it as a new input. The original plan remains available as the baseline. Mutating a saved plan in place would make it harder to establish whether an unexpected difference came from the requested edit or from accidental artifact drift. A separate result directory for each attempt also keeps successful and failed versions connected to the request that produced them.

The contract should also specify what will count as evidence. A symbolic comparison can test whether declared score events remained fixed. A result record can show which request, model, score, and decoder were used. Listening can reveal whether the intended harmony is audible and whether the transition into and out of a changed passage works. None of these checks substitutes for the others.